Computational Models For The Human Body PDF

General Preface
In the early eighties, when Jacques-Louis Lions and I considered the idea of a Hand-
book of Numerical Analysis, we carefully laid out specific objectives, outlined in the
following excerpts from the “General Preface” which has appeared at the beginning of
each of the volumes published so far:
During the past decades, giant needs for ever more sophisticated mathe-
matical models and increasingly complex and extensive computer simula-
tions have arisen. In this fashion, two indissociable activities, mathematical
modeling and computer simulation, have gained a major status in all aspects
of science, technology and industry.
In order that these two sciences be established on the safest possible
grounds, mathematical rigor is indispensable. For this reason, two compan-
ion sciences, Numerical Analysis and Scientific Software, have emerged as
essential steps for validating the mathematical models and the computer
simulations that are based on them.
Numerical Analysis is here understood as the part of Mathematics that de-
scribes and analyzes all the numerical schemes that are used on computers;
its objective consists in obtaining a clear, precise, and faithful, representa-
tion of all the “information” contained in a mathematical model; as such, it
is the natural extension of more classical tools, such as analytic solutions,
special transforms, functional analysis, as well as stability and asymptotic
analysis.
The various volumes comprising the Handbook of Numerical Analysis
will thoroughly cover all the major aspects of Numerical Analysis, by pre-
senting accessible and in-depth surveys, which include the most recent
trends.
More precisely, the Handbook will cover the basic methods of Numerical
Analysis, gathered under the following general headings:
– Solution of Equations in Rn ,
– Finite Difference Methods,
– Finite Element Methods,
– Techniques of Scientific Computing.
v
vi General Preface
It will also cover the numerical solution of actual problems of contempo-

rary interest in Applied Mathematics, gathered under the following general
headings:
– Numerical Methods for Fluids,
– Numerical Methods for Solids.
In retrospect, it can be safely asserted that Volumes I to IX, which were edited by
both of us, fulfilled most of these objectives, thanks to the eminence of the authors and
the quality of their contributions.
After Jacques-Louis Lions’ tragic loss in 2001, it became clear that Volume IX would
be the last one of the type published so far, i.e., edited by both of us and devoted to some
of the general headings defined above. It was then decided, in consultation with the pub-
lisher, that each future volume will instead be devoted to a single “specific application”
and called for this reason a “Special Volume”. “Specific applications” will include Math-
ematical Finance, Meteorology, Celestial Mechanics, Computational Chemistry, Living
Systems, Electromagnetism, Computational Mathematics etc. It is worth noting that the
inclusion of such “specific applications” in the Handbook of Numerical Analysis was
part of our initial project.
To ensure the continuity of this enterprise, I will continue to act as Editor of each Spe-
cial Volume, whose conception will be jointly coordinated and supervised by a Guest
Editor.
P.G. C IARLET
July 2002
Foreword
Computational Models for the Human Body constitute an emerging and rapidly pro-
gressing area of research whose primary objective is to provide a better understanding
of the physiological and mechanical behavior of the human body and to design tools for
their realistic numerical simulations. This volume describes concrete examples of such
computational models. Although far from being exhaustive, it covers a large range of
methods and an illustrative set of applications, and proposes a number of well-defined
mathematical and numerical modeling of physical problems (including the analysis of
existence and uniqueness of solutions for instance), followed by various numerical sim-
ulations.
Medical applications are addressed first, because physiological and biomechanical
models of the human body already play a prominent role in the prevention, diagnosis
and therapy of many diseases. The generalized introduction of such models in medicine
will in fact strongly contribute to the development of a more individualized and preven-
tive medicine. In effect, through the continuous progress of medical imaging during the
past decades, it is currently possible to extract an increasing flow of anatomical or func-
tional information on any individual, with an increasingly accurate resolution in space
and time. The overwhelming quantity of available signals and images makes a direct
analysis of the data more and more difficult, when not impossible. New computational
models are necessary to capture those parameters that are pertinent to analyze the human
system under study or to simulate it. There is also a number of important non-medical
applications of these computational models which cover numerous human activities,
like driving (safer design of vehicles), working (better ergonomy of workplaces), exer-
cising (more efficient training of athletes), entertaining (simulation for movies), etc.
There are basically three levels of design for human models. The first level is mainly
geometrical and addresses the construction of a digital description of the anatomy, of-
ten acquired from medical imagery. The second level is physical, involving mainly the
biomechanical modeling of various tissues, organs, vessels, muscles or bone structures.
The third level is physiological, involving a modeling of the functions of the major bio-
logical systems (e.g., cardiovascular, respiratory, digestive, hormonal, muscular, central
or peripheral nervous system, etc.) or some pathological metabolism (e.g., evolution of
cancerous or inflammatory lesions, formation of vessel stenoses, etc.). A fourth level
(not described in this volume) would be cognitive, modeling the higher functions of the
human brain. These different levels of modeling are closely related to each other, and
vii
viii Foreword
several physiological systems may interact together (e.g., the cardiopulmonary interac-
tion). The choice of the resolution at which each level is described is important, and
may vary from microscopic to macroscopic, ideally through multiscale descriptions.
The first three chapters of this volume study three important physiological models
(vascular, cardiac, and tumoral) from a mathematical and numerical perspective. The
chapter by Alfio Quarteroni and Luca Formaggia addresses the problem of developing
models for the numerical simulation of the human circulatory system, focussing on the
analysis of haemodynamics in arteries. Applications include the prediction (and there-
fore the possible prevention) of stenoses (a local reduction of the lumen of the artery), a
leading cause of cardiovascular accidents. The chapter by Mary Belik, Taras Usyk and
Andrew McCulloch describes computational methods for modeling and simulating the
cardiac electromechanical function. These methods provide tools to predict physiologi-
cal function from quantitative measurements of tissue, cellular or molecular structures.
Applications include a better understanding of cardiac pathologies, and a quantitative
modeling of their evolution from various sources of measurements, including medical
imagery. The chapter by Jesús Ildefonso Díaz and José Ignacio Tello studies the mathe-
matical properties of a simple model of tumor growth. Proofs are given for the existence
and uniqueness of solutions and numerical simulations of the model are presented.
The next two chapters are dedicated to the simulation of deformations inside the
human body in two different contexts. The chapter by Eberhard Haug, Hyung-Yun Choi,
Stéphane Robin and Muriel Beaugonin describes computational models for crash and
impact simulation. It presents the latest generation of virtual human models used to
study the consequences of car accidents on organs and important anatomical structures.
These models allow the interactive design of safer vehicles with an unrivaled flexibility.
The chapter by Hervé Delingette and Nicholas Ayache describes computational models
of soft tissue useful for surgery simulation. The real-time constraint imposed by the
necessary realism of a training system leads to specific models which are applied to the
simulation of minimally invasive digestive surgery, including liver surgery.
The last two chapters describe computational models dedicated to image-guided in-
tervention and diagnosis. The chapter by Xenophon Papademetris, Oskar Skrinjar and
James Duncan describes computational models of organs used to predict and track de-
formations of tissues from sparse information acquired through medical imaging. These
models rest on a successful combination of biomechanical modeling with medical im-
age analysis, with an application to image-guided neurosurgery and an application to the
image-based quantitative analysis of cardiac diseases. The chapter by Fred Azar, Dim-
itris Metaxas and Mitchell Schnall presents a computational model of the breast used to
predict deformations during interventions. The main applications are for image-guided
clinical biopsies and for image-guided therapy.
Before concluding this introduction, I wish to wholeheartedly thank all the authors
for their essential contributions, their patience and confidence during all the genesis
process of this book. Special thanks are due to my colleague Hervé Delingette, whose
advice was extremely helpful from the very beginning. I wish to thank several col-
leagues for their important help and the many improvements they suggested: Michel
Audette, Chris Berenbruch, Mark Chaplain, Olivier Clatz, Stéphane Lanteri, Denis Lau-
rendeau, Philippe Meseure, Serge Piperno, Jean-Marc Schwartz, Brian Sleeman, Michel
Foreword ix
Sorine, Matthias Teschner, Marc Thiriet, Marina Vidrascu. I also wish to thank Gilles
Kahn, Scientific Director of INRIA, who has been extremely supportive of this project
originating from our institute.
Finally, I wish to honor the memory of Jacques-Louis Lions, who contacted me for
the first time at the end of November 1999 with the proposition to work on this project.
The original title changed several times, before finally converging towards its final title
after recent discussions with Philippe Ciarlet, to whom will go my final thanks, for his
great encouragements and confidence.
N ICHOLAS AYACHE
Sophia–Antipolis, France
1st November 2003
This page intentionally left blank
Contents of Volume XII
S PECIAL VOLUME : C OMPUTATIONAL M ODELS FOR THE H UMAN B ODY
G ENERAL P REFACE v
F OREWORD vii
Mathematical Modelling and Numerical Simulation of the Cardiovascular
System, A. Quarteroni, L. Formaggia 3
Computational Methods for Cardiac Electrophysiology, M.E. Belik,
T.P. Usyk, A.D. McCulloch 129
Mathematical Analysis, Controllability and Numerical Simulation of a
Simple Model of Avascular Tumor Growth, J.I. Díaz, J.I. Tello 189
Human Models for Crash and Impact Simulation, E. Haug, H.-Y. Choi,
S. Robin, M. Beaugonin 231
Soft Tissue Modeling for Surgery Simulation, H. Delingette, N. Ayache 453
Recovering Displacements and Deformations from 3D Medical Images
Using Biomechanical Models, X. Papademetris, O. Škrinjar, J.S. Duncan 551
Methods for Modeling and Predicting Mechanical Deformations of the
Breast under External Perturbations, F.S. Azar, D.N. Metaxas,
M.D. Schnall 591
S UBJECT I NDEX 657
xi
Mathematical Modelling
and Numerical Simulation
of the Cardiovascular
System
Alfio Quarteroni a,b, Luca Formaggia b

a Institute of Mathematics, EPFL, Lausanne, Switzerland
b MOX, Department of Mathematics, Politecnico di Milano, Milano, Italy
E-mail addresses: alfio.quarteroni@epfl.ch (A. Quarteroni),

luca.formaggia@mate.polimi.it (L. Formaggia)
Computational Models for the Human Body Copyright © 2004 Elsevier B.V.
Special Volume (N. Ayache, Guest Editor) of All rights reserved
HANDBOOK OF NUMERICAL ANALYSIS, VOL. XII ISSN 1570-8659
P.G. Ciarlet (Editor) DOI 10.1016/S1570-8659(03)12001-7
3
Contents
C HAPTER I 7
1. Introduction 7
2. A brief description of the human vascular system 8
3. The main variables for the mathematical description of blood flow 11
4. Some relevant issues 11
C HAPTER II 17
5. The derivation of the equations for the flow field 17

6. Some nomenclature 17
7. The motion of continuous media 18
8. The derivation of the basic equations of fluid mechanics 25
9. The Navier–Stokes equations 31
C HAPTER III 35
10. The incompressible Navier–Stokes equations and their approximation 35

11. Weak form of Navier–Stokes equations 37
12. An energy inequality for the Navier–Stokes equations 41
13. The Stokes equations 43
14. Numerical approximation of Navier–Stokes equations 44
C HAPTER IV 57
15. Mathematical modelling of the vessel wall 57

16. Derivation of 1D models of vessel wall mechanics 59
17. Analysis of vessel wall models 68
C HAPTER V 71
18. The coupled fluid structure problem 71

19. An iterative algorithm to solve the coupled fluid–structure problem 80
C HAPTER VI 87
20. One-dimensional models of blood flow in arteries 87
5
6 A. Quarteroni and L. Formaggia
C HAPTER VII 111
21. Some numerical results 111

22. Conclusions 122
Acknowledgements 122
R EFERENCES 125
C HAPTER I
1. Introduction
In these notes we will address the problem of developing models for the numerical
simulation of the human circulatory system. In particular, we will focus our attention
on the problem of haemodynamics in large human arteries.
Indeed, the mathematical investigation of blood flow in the human circulatory system
is certainly one of the major challenges of the next years. The social and economical
relevance of these studies is highlighted by the unfortunate fact that cardiovascular dis-
eases represent the major cause of death in developed countries.
Altered flow conditions, such as separation, flow reversal, low and oscillatory shear
stress areas, are now recognised by the medical research community as important factors
in the development of arterial diseases. An understanding of the local haemodynamics
can then have useful applications for the medical research and, in a longer term per-
spective, to surgical planning and therapy. The development of effective and accurate
numerical simulation tools could play a crucial role in this process.
Besides their possible role in medical research, another possible use of numerical
models of vascular flow is to form the basis for simulators to be used as training sys-
tems. For instance, a technique now currently used to cure a stenosis (a pathological
restriction of an artery, usually due to fat deposition) is angioplasty. It consists of in-
flating a balloon positioned in the stenotic region by the help of a catheter. The balloon
should squash the stenosis and approximately restore the original lumen area. The suc-
cess of the procedure depends, among other things, on the sensitivity of the surgeon and
his ability of placing the catheter in the right position. A training system which couples
virtual reality techniques with the simulation of the flow field around the catheter, the
balloon and the vessel walls, employing geometries extracted from real patients, could
well serve as training bed for new vascular surgeons. A similar perspective could pro-
vide specific design indications concerning the realisations of surgical operations. For
instance, numerical simulations could help the surgeon in understanding how the differ-
ent surgical solutions may affect blood circulation and guide the selection of the most
appropriate procedure for a specific patient.
In such “virtual surgery” environments, the outcome of alternative treatment plans for
the individual patient can be foreseen by simulations. This numerical approach is one
of the aspects of a new paradigm of the clinical practice, which is referred to as “predic-
tive medicine” (see TAYLOR , D RANEY, K U , PARKER , S TEELE , WANG and Z ARINS
[1999]).
Since blood flow interacts mechanically with the vessel walls, it gives rise to a rather
complex fluid–structure interaction problem which requires algorithms able to correctly
7
8 A. Quarteroni and L. Formaggia C HAPTER I
describe the energy transfer between the fluid (typically modelled by the Navier–Stokes
equations) and the structure. This is indeed one of the main subjects of these notes,
which will adopt the following steps:
(1) Analysis of the physical problem. We illustrate problems related to haemody-
namics, focusing on those aspects which are more relevant to human physiology.
This will allow us to identify the major mathematical variables useful for our
investigation. This part will be covered in Section 2.
(2) Mathematical modelling. Starting from some basic physical principles, we will
derive the partial differential equations which link the variables relevant to the
problem. We will address some difficulties associated to the specific characteris-
tics of these equations. Problems such as existence, uniqueness and data depen-
dence of the solution will be briefly analysed. In particular, in Section 5 we will
deal with models for the fluid flow and recall the derivation of the incompress-
ible Navier–Stokes equations starting from the basic principles of conservation of
mass and momentum. In Section 15 the attention will be instead focused on the
dynamics of the vessel wall structure. Some simple, yet effective, mathematical
models for the vessel wall displacement will be derived and discussed.
(3) Numerical modelling. We present different schemes which can be employed to
solve the equations that have been derived and discuss their properties. In par-
ticular, Section 10 deals with some relevant mathematical aspects related to the
numerical solution of the equations governing the flow field, while Section 18 is
dedicated to the coupled fluid–structure problem.
Reduced models which make use of a one dimensional description of blood
flow in arteries are often used to study the propagation of average pressure and
mass flow on segments of the arterial tree. In Section 20 we present the derivation
of a model of this type, together with a brief analysis of its main mathematical
characteristics.
(4) Numerical simulation. A final section is dedicated to numerical results obtained
on relevant test cases.
2. A brief description of the human vascular system
The major components of the cardiovascular system are the heart, the arteries and the
veins. It is usually subdivided into two main parts: the large circulation system and
the small circulation system, as shown in Fig. 2.1. The former brings oxygenated blood
from the heart left ventricle to the various organs (arterial system) and then brings it back
to right atrium (venous system). The latter pumps the venous blood into the pulmonary
artery, where it enters the pulmonary system, get oxygenated and is finally received by
the heart left atrium, ready to be sent to the large circulation system.
Fig. 2.2 shows a picture of the human heart. Its functioning is very complex and
various research teams are currently trying to develop satisfactory mathematical models
of its mechanics, which involves, among other things, the study of the electro-chemical
activation of the muscle cells. We will not cover this aspect in these notes, where we
rather concentrate on vascular flow and, in particular, flow in arteries.
S ECTION 2 9
F IG . 2.1. The human circulatory system. The human cardiovascular system has the task of supplying the
human organs with blood. Its correct working is obviously crucial and depends on many parameters: external
temperature, muscular activity, state of health, just to mention a few. The blood pressure and flow rate then
change according to the body needs.
F IG . 2.2. The human heart. Courtesy of the Texas Heart®Institute.

Arteries can be regarded as hollow tubes with strongly variable diameters and can
be subdivided into large arteries, medium arteries and arterioles and capillaries. The
main role of large arteries (1–3 cm of diameter) is to carry a substantial blood flow
rate from the heart to the periphery and to act as a “compliant system”. They deform
under blood pressure and by doing so they are capable of storing elastic energy during
the systolic phase and return it during the diastolic phase. As a result the blood flow
is more regular than it would be if the large arteries were rigid. We then have a fluid–
structure interaction problem. The blood may be considered a homogeneous fluid, with
“standard” behaviour (Newtonian fluid), the wall may be considered elastic (or mildly
visco-elastic).
The smaller arteries (0.2 mm–1 cm of diameter) are characterised by a strong branch-
ing. The vessel may in general be considered rigid (apart in the heart, where the vessel
movement is mainly determined by the heart motion). Yet, the blood begins to show
“non-standard” behaviour typical of a shear-thinning (non-Newtonian) fluid.
The arterioles have an important muscular activity, which is aimed at regulating blood
flow to the periphery. Consequently, the vessel wall mechanical characteristics may
change depending on parameters such as blood pressure and others. At the smallest
levels (capillaries), blood cannot be modelled anymore as a homogeneous fluid, as the
dimension of the particles are now of the same order of that of the vessel. Furthermore,
the effect of wall permeability on the blood flow becomes important.
The previous subdivision is not a mere taxonomy: the morphology of the vessel
walls and the physical characteristics of blood change in dependence of the type of
vessel.
Indeed, the blood is not a fluid but a suspension of particles in a fluid called plasma.
Blood particles must be taken into account in the rheological model in smaller arterioles
and capillaries since their size becomes comparable to that of the vessel. The most
important blood particles are:
• red cells (erythrocytes), responsible for the exchange of oxygen and carbon-
dioxide with the cells;
• white cells (leukocytes), which play a major role in the human immune system;
• platelets (thrombocytes), main responsible for blood coagulation.
Here, we will limit to flow in large/medium sized vessels. We have mentioned that
the vascular system is highly complex and able to regulate itself: an excessive decrease
in blood pressure will cause the smaller arteries (arterioles) to contract and the heart
rate increase. On the contrary, an excessive blood pressure is counter-reacted by a re-
laxation of the arterioles wall (which causes a reduction of the periphery resistance to
the flow) and decreasing the heart beat. Yet, it may happen that some pathological con-
ditions develop, for example, the arterial wall may become more rigid, due to illness or
excessive smoking habits, fat may accumulate in some areas causing a stenosis, that is
a reduction of the vessel section as illustrated in Fig. 2.3, aneurysms may develop. The
consequence of these pathologies on the blood field as well as the possible outcome of
a surgical intervention may be studied by numerical tools.
S ECTION 3 11
F IG . 2.3. The deposition of lipids and cholesterol in the inner wall of an artery (frequently a coronary) can
cause a stenosis and eventually a dramatic reduction (or even the interruption) of blood flow. Images taken
from “Life: the Science of Biology” by W.K. Purves et al., fourth edition, published by Sinauer Associates
Inc. and W.H. Freeman and Company.
3. The main variables for the mathematical description of blood flow
The principal quantities which describe blood flow are the velocity u and pressure P .
Knowing these fields allows the computation of the stresses to which an arterial wall
is subjected due to the blood movement. Since we will treat fluid–structure interaction
problems, the displacement of the vessel wall due to the action of the flow field is
another quantity of relevance. Pressure, velocity and vessel wall displacement will be
functions of time and the spatial position.
The knowledge of the temperature field may also be relevant in some particular con-
text, such as the hyperthermia treatment, where some drugs are activated through an
artificial localised increase in temperature. Temperature may also have a notable influ-
ence on blood properties, in particular on blood viscosity. Yet, this aspect is relevant
only in the flow through very small arterioles/veins and in the capillaries, a subject
which is not covered in these notes.
Another aspect of blood flow which we will not cover in these notes, is the chem-
ical interaction with the vessel wall, which is relevant both for the physiology of the
blood vessels and for the development of certain vascular diseases. Not mentioning the
potential relevance of such investigation for the study of the propagation/absorption
of pharmaceutical chemicals. Some numerical models and numerical studies for the
chemical transport/diffusion process in blood and through arterial wall may be found in
R APPITSCH and P ERKTOLD [1996], Q UARTERONI , V ENEZIANI and Z UNINO [2002].
4. Some relevant issues
Among the difficulties in the modelling of blood flow in large vessels, we mention the
following ones:
F IG . 4.1. A typical flow rate in an artery during the cardiac cycle.
• The flow is transient. Blood flow is obviously pulsatile. This means that one can-
not neglect the time by considering a “steady state” solution, function only of the
spatial position, as it is often done in many other situations (for example, the study
of the flow field around an aeroplane or a car). With some approximation one
may think the blood flow to be periodic in time. Yet, this is usually true only for
relatively short periods, since the various human activities require to change the
amount of blood sent to the various organs.
The cardiac cycle can be subdivided into two phases. The systole corresponds
to the instant in which the heart is pumping the blood into the arterial system. The
systolic period is then characterised by the highest flow rate. The diastole, instead,
corresponds to the instant in which the heart is filling up with the blood coming
from the venous system and the aortic valve is closed. The blood flow is then at its
minimum. Fig. 4.1 illustrates a typical flow rate curve on a large artery during the
cardiac cycle.
Unsteady flow is usually much more complex than its steady counterpart. For
instance, if we consider a steady flow of a fluid like water inside an “infinitely
long” cylindrical tube, it is possible to derive the analytical steady state solution
(also called the Poiseuille flow solution), characterised by a parabolic velocity pro-
file. Transient flow in the same geometrical configuration becomes much more
complex. The solution may still be obtained analytically if we assume time peri-
odicity, giving rise to the so-called Womersley flow (W OMERSLEY [1955]), whose
expression may be found, for instance, in Q UARTERONI , T UVERI and V ENEZIANI
[2000]. Just as an example, in Fig. 4.2 we show the velocity profile in a tube for a
Poiseuille and for a Womersley flow (the latter, obviously, at a given instant) (from
V ENEZIANI [1998]).
• The wall interacts mechanically with the flow field. This aspect is relevant for rel-
atively large vessels. In the aorta, for example, the radius may vary in a range of
5 to 10% between diastole and systole. This is quite a large displacement, which
affects the flow field. The fluid–structure interaction problem is the responsible for
S ECTION 4 13
F IG . 4.2. Three-dimensional velocity profiles for a Poiseuille flow (left) and Womersley unsteady flow at
a given instant (right).
the propagation of pulse pressure waves. Indeed, no propagative phenomena would

otherwise occur in an incompressible fluid like blood. The interaction problem is a
rather complex one, since the time scales associated to the interaction phenomena
are two orders of magnitude greater than those associated to the bulk flow field.
In arterioles and capillaries the movement of the wall may be considered negli-
gible.
• Lack of boundary data. We are normally interested in modelling only a section
of the cardiovascular system by means of partial differential equations. A proper
setting of a differential problem requires to provide appropriate conditions at the
domain boundary, i.e., on the sections at the ends of the region of interest. For
instance, let us consider Fig. 4.3. “Standard” conditions for the inlet section Γ in
and the outlet sections Γ out , may be derived from the analysis of the differential
equations governing the fluid flow. A possible choice is to prescribe all compo-
nents of the velocity on Γ in and the velocity derivative along the normal direction
F IG . 4.3. An example of a computational domain made of a section of vascular system. We need to provide
proper boundary conditions at Γ in , Γ out and Γ w .
(or the normal stress components) on Γ out . Unfortunately, in practise one never
has enough data for prescribing all these conditions. Normally, only “averaged”
data are available (mean velocity and mean pressure), which are not sufficient for
a “standard” treatment of the mathematical problem. One has thus to devise al-
ternative formulations for the boundary conditions which, on one hand reflect the
physics and exploit the available data, on the other hand, permit to formulate a
mathematically well posed problem. In these notes we will not investigate this par-
ticular aspect. A possible formulation for the flow boundary conditions which is
particularly suited for vascular flow problems is illustrated and analysed in F OR -
MAGGIA , G ERBEAU , N OBILE and Q UARTERONI [2002].
We have not used the terms “inflow” and “outflow” to indicate boundary conditions
at Γ in and Γ out since they would be incorrect. Indeed, outflow would indicate the nor-
mal component of the velocity is everywhere positive (while it is negative at an inflow
section). However, in vascular problems, this assumption is seldom true because the
pulsating nature of blood flow might (and typically does) induce a flow reversal on
portions of an artery during the cardiac beat. Indeed, the Womersley solution (W OM -
ERSLEY [1955]) of a pulsatile flow in circular cylinders, which provides a reasonable
approximation of the general flow pattern encountered in arteries, shows a periodic flow
reversal.
In the medical literature, one encounters the terms “proximal” to indicate the section
which is reached first by the flow exiting from the heart, while “distal” is the term
associated to the sections which are farther from the heart. Here we have preferred
instead the terms “inlet” and “outlet” which refer to the behaviour of the mean flow rate
across the section. At an inlet (outlet) section the mean flow is entering (exiting) the
vascular element under consideration.
Some of the problems which the simulation of blood flow in large arteries may help
in answering are summarised below.
F IG . 4.4. Recirculation in the carotid bifurcation. On the left we illustrate the location of the carotid bifurca-
tion. The image on the right shows the particle path during the diastolic period in a model of the carotid bifur-
cation. A strong recirculation occurs inside the carotid sinus. The image on the left is courtesy of vesalius.com.
S ECTION 4 15
F IG . 4.5. A schematic example of a coronary by-pass. The alteration of the flow field due to the by-pass may
cause the formation of a new stenosis, typically immediately downstream the by-pass.
• Study of the physiological behaviour of vessel walls. For example, are there any
characteristics of the flow field which may be related to the formation of stenoses?
In particular, in some sites like the carotid bifurcation (see Fig. 4.4) it is quite
usual to have a reversal of the flow during the cardiac cycle which generates a
recirculation zone. These recirculation zones have been found to be possible sites
for fat accumulation and, consequently, the appearance of stenosis. There is some
evidence that one of the factors which prompt fat accumulation is linked to the
oscillatory nature of the vessel wall stresses induced by the fluid in the flow reversal
zone. Wall stresses are quantities very difficult to measure “in vivo” while are
easily computed once the flow field is known. Numerical simulations may then
help in assessing the effectiveness of such theory.
• Study of post-surgical situations. Is it possible to predict the flow behaviour af-
ter the geometry has been modified by a surgical operation like a by-pass (see
Fig. 4.5)? It has been found that the flow pattern in the by-pass region may af-
fect the insurgence of post-surgery pathologies. Again, a zone with recirculating
or stagnant fluid has negative consequences. Numerical simulations may allow to
predict the post-surgery flow pattern and determine, say, the best by-pass configu-
ration.
C HAPTER II
5. The derivation of the equations for the flow field
The flow field is governed by a set of partial differential equations in a region whose
boundary changes in time. Their derivation, moving from the basic physical principles
of conservation of mass and momentum, is the scope of this chapter.
6. Some nomenclature
The space R3 is equipped with a Cartesian coordinate system defined by the orthonor-
mal basis (e1 , e2 , e3 ), where

1 0 0
e1 = 0 , e2 = 1 , e3 = 0 .
0 0 1
Vectors are understood as column vectors. A vector f ∈ R3 may then be written as

3
f= fi ei ,
i=1
where fi is the ith component of f with respect to the chosen basis. Vectors will be
always indicated using bold letters while their components will be generally denoted
by the same letter in normal typeface. Sometimes, when necessary for clarity, we will
indicate the ith component of a vector f by (f)i or simply fi . These definitions apply
to vectors in R2 as well. With the term domain we will indicate an open, bounded,
connected subset of RN , N = 2, 3, with orientable boundary. We will indicate with n
the outwardly oriented unit vector normal to the boundary. We will also assume that
the domain boundary be Lipschitz continuous (for instance, a piece-wise polynomial,
or a C 1 curve). In Fig. 6.1 some admissible domains are shown. If a quantity f (like
temperature or pressure) takes a scalar value on a domain Ω, we say that the quantity
defines a scalar field on Ω, which we will indicate with f : Ω → R. If instead a quantity
f associates to each point in Ω a vector (as in the case of the velocity), we say that
it defines a vector field on Ω, and we will indicate it with f : Ω → R3 . Finally, if a
quantity T associates to each point in Ω a RN×N matrix, we will say that it defines a
(second order) tensor field on Ω if it obeys the ordinary transformation rules for tensors
(A RIS [1962]). Its components will be indicated by either (T)ij , or simply Tij , with
i, j = 1, . . . , 3.
17
18 A. Quarteroni and L. Formaggia C HAPTER II
F IG . 6.1. Example of admissible domains. Ωa has a boundary formed by piece-wise C 1 curves. Ωb is a

multi-connected domain, with a polygonal internal and a C ∞ external boundary. Finally, Ωc has a polygonal
boundary.
Given a function f : Ω → R, x → f (x), and a domain V ⊂ Ω, we will use the

shorthand notation

f
V
to indicate the integral

f (x) dx,
V
and

f
∂V
to indicate the surface (or line) integral

f dσ,
∂V
unless the context requires otherwise.
When referring to a physical quantity f , we will indicate with [f ] its measure units
(in the international system). For instance, if v indicates a velocity, [v] = m/s, where m
stands for meters and s for seconds.
7. The motion of continuous media
In order to derive the differential equations which govern the fluid motion, we need
to introduce some kinematic concepts and quantities. The kinematics of a continuous
medium studies the property of the motion of a medium which may be thought as con-
tinuously occupying, at each time, a portion of space. This allows the use of standard
methods of analysis. We will set the derivation in R3 , since this is the natural spatial
S ECTION 7 19
F IG . 7.1. The Lagrangian mapping.
dimension. However, the definitions and final differential equations are valid also in R2 .
Furthermore, we will assume that the motion will take place during a time interval
I = (t0 , t1 ).
The motion itself is described by a family of mappings Lt which associate the posi-
tion x of a fluid particle at time t ∈ I to a point ξ ∈ Ω0 , Ω0 being the domain occupied
by the fluid at the reference initial time t0 . More precisely, we denote with Ωt the portion
of space occupied by the fluid at time t and we indicate with Lt the mapping
Lt : Ω0 → Ωt , ξ → x = x(t, ξ ) = Lt (ξ ),
which will be denoted Lagrangian mapping at time t (see Fig. 7.1). We assume that Lt
is continuous and invertible in Ω 0 , with continuous inverse.
We call Ω0 the reference configuration, while Ωt is called current (or spatial) con-
figuration. The position of the material particle located at the point x in the current
configuration Ωt is a function of time and of the position of the same material particle
at the reference time.
We may thus relate the variables (t, x) to (t, ξ ). The former couple is referred to as
the Eulerian variables while the latter are called the Lagrangian variables.
It is worthwhile to point out that when using the Eulerian variables as independent
variables, we are concentrating our attention on a position in space x ∈ Ωt and on the
fluid particle which, at that particular time, is located at x. When using the Lagrangian
variables as independent variables (Lagrangian frame), we are instead targeting the fluid
particle “labelled” ξ (that is the fluid particle which was located at position ξ at the
reference time). That is, we are following the trajectory Tξ of fluid particle ξ ∈ Ω0 ,
defined as

Tξ = t, x(t, ξ) : t ∈ I . (7.1)
The basic principles of mechanics are more easily formulated with reference to the
moving particles, thus in the Lagrangian frame. Yet, in practice it is more convenient to
work with the Eulerian variables. Therefore, we need to rewrite the equations stemming
from those basic principles into the Eulerian frame. We will see later on that for the
numerical approximation of the problem at hand it will be necessary to introduce yet

another, intermediate, frame of reference, called Arbitrary Lagrangian Eulerian.
Being the mapping surjective, a quantity associated with the fluid may be described as
function of either the Lagrangian or the Eulerian variables, depending on convenience.
We will in general use the same symbol for the functions which describe the evolution
of the same quantity in the Lagrangian and in the Eulerian frame, unless the context
needs otherwise. In the latter case, we will mark with the hat symbol “ ˆ ” a quantity
expressed as function of the Lagrangian variables, that is, if f : I × Ωt → R we have
the equality
fˆ(t, ξ ) = f (t, x), with x = Lt (ξ ).
We will often use the following alternative notation:
fˆ = f ◦ Lt or f = fˆ ◦ L−1
t
with the understanding that the composition operator applies only to the spatial vari-
ables.
The symbol ∇ is used exclusively to indicate the gradient with respect to the Eulerian
variable x. When we need to indicate the gradient with respect to the Lagrangian vari-
able ξ , we will use the symbol ∇ ξ , that is

3
∂ fˆ
∇ ξ fˆ = ei .
∂ξi
i=1
The same convention applies to other spatial differential operators (divergence, Lapla-
cian, etc.) as well.
In the following we will put I × Ωt = {(t, x): t ∈ I, x ∈ Ωt } (note the little abuse of
notation since technically it is not a Cartesian product).
7.1. The velocity
The fluid velocity is the major kinematic quantity of our problem. In the Lagrangian
frame it is expressed by means of a vector field û = û(t, ξ ) defined as
∂x ∂
û = , i.e., û(t, ξ ) = x(t, ξ). (7.2)
∂t ∂t
û is called the Lagrangian velocity field (or velocity field in the Lagrangian frame), and
it denotes the time derivative along the trajectory Tξ of the fluid particle ξ . The velocity
u on the Eulerian frame is defined for (t, x) ∈ I × Ωt as

u = û ◦ L−1 −1
t , i.e., u(t, x) = û t, Lt (x) .
E XAMPLE 7.1. Let us consider a 2D case and the following movement law, for t 0:
x1 = ξ1 et , ξ1 ∈ (−1, 1), x2 = ξ2 , ξ2 ∈ (−1, 1).
The domain at time t > 0 occupies the rectangle (−et , et ) × (−1, 1). The mapping is
clearly invertible for all t 0.
S ECTION 7 21
We have
û1 = ∂x1 /∂t = ξ1 et , û2 = ∂x2 /∂t = 0.
We can immediately compute the velocity field as function of the Eulerian variable as
u1 = x 1 , u2 = 0.
Once the velocity field and the reference configuration is known, the motion may be
derived by solving the following Cauchy problem:
For any ξ ∈ Ω0 , find the function x = x(t, ξ ) which satisfies

 ∂x
(t, ξ ) = û(t, ξ ), ∀t ∈ I,
∂t

x(t0 , ξ ) = ξ .
7.2. The material derivative
We can relate time derivatives computed with respect to the different frames. The mate-
rial (or Lagrangian) time derivative of a function f , which we will denote Df /Dt , is
defined as the time derivative in the Lagrangian frame, yet expressed as function of the
Eulerian variables.
That is, if f : I × Ωt → R and fˆ = f ◦ Lt ,
Df Df ∂ fˆ
: I × Ωt → R, (t, x) = (t, ξ ), ξ = L−1
t (x). (7.3)
Dt Dt ∂t
Therefore, for any fixed ξ ∈ Ω0 we may also write
Df d
(t, x) = f t, x(t, ξ ) ,
Dt dt
by which we can observe that the material derivative represents the rate of variation of
f along the trajectory Tξ .
By applying the chain-rule of derivation of composed functions, we have
Df ∂f
= + u ·∇ f. (7.4)
Dt ∂t
Indeed,

Df ∂ −1 ∂f ∂x −1 ∂f
= (f ◦ Lt ) ◦ Lt = + ∇f · ◦ Lt = + u · ∇f.
Dt ∂t ∂t ∂t ∂t
A quantity which satisfies
∂f
=0
∂t
is called stationary, and a motion for which
∂u
=0
∂t
is said a stationary motion.
E XAMPLE 7.2. Let us consider again the motion of Example 7.1 and consider the func-
tion f (x1 , x2 ) = 3x1 + x2 (which is independent of t). The application of relation (7.4)
gives

Df x 3
=0+ 1 · = 3x1 .
Dt 0 1
On the other hand,
∂ fˆ
fˆ = 3ξ1 et + ξ2 and = 3ξ1 et ,
∂t
by which we deduce that,
∂ fˆ
◦ L−1
t = 3x1 .
∂t
This example, besides verifying relation (7.4), shows that a function f = f (t, x) with
∂f /∂t = 0 in general has Df/Dt = 0.
7.3. The acceleration
In the Lagrangian frame the acceleration is a vector field â : I × Ω0 → R3 defined as
∂ û ∂ 2 x
â = = 2.
∂t ∂t
By recalling the definition of material derivative, we may write the acceleration in
Eulerian frame as
Du ∂u
a= = + (u · ∇)u. (7.5)
Dt ∂t
Component-wise,
∂ui ∂ui
3
ai = + uj . (7.6)
∂t ∂xj
j =1
7.4. The deformation gradient
Another kinematic quantity necessary for the derivation of the mathematical model is
the deformation gradient
Ft , which is defined, for each t ∈ I , as
∂x
Ft : Ω0 → RN×N , Ft = ∇ m Lt = . (7.7)
∂ξ
Component-wise,
∂xi
(
Ft )ij = .
∂ξ j
S ECTION 7 23
In particular, its determinant,

Jt = det
Ft , (7.8)
is called the Jacobian of the mapping Lt . As usual, its counterpart in the Eulerian frame
is indicated Jt .
It is possible to show that the time continuity and the invertibility of the Lagrangian
mapping is sufficient to have, for all t ∈ I ,
Jt (ξ ) > 0 ∀ξ ∈ Ω0 . (7.9)
The importance of Jt is clearly linked to the rule which transforms integrals from the
current to the reference configuration. We recall the following theorem of elementary
calculus (without providing its proof).
T HEOREM 7.1. Let Vt ⊂ Ωt be a subdomain of Ωt and let us consider the function

f : I × Vt → R. Then f is integrable on Vt if and only if (f ◦ Lt )Jt is integrable on
V0 = L−1t (Vt ), and

f (t, x) dx = fˆ(t, ξ ) Jt (ξ ) dξ ,
Vt V0
where fˆ(t, ξ ) = f (t, Lt (ξ )). In short,

f= fˆJt .
Vt V0
7.5. The Reynolds transport theorem
An interesting property of the Jacobian is that its time derivative is linked to the diver-
gence of the velocity field.
L EMMA 7.1. Let Jt denote the Jacobian (7.8) in the Eulerian frame. Then
D
Jt = Jt div u. (7.10)
Dt
This relation is sometimes called Euler expansion formula.
P ROOF. We have, by direct application of the chain-rule,

∇ ξ û = ∇ ξ (u ◦ Lt ) = ∇u
∇ ξ Lt = ∇u Ft .
On the other hand, by recalling the definition of the velocity (7.2),

∂x ∂ ∂
Ft
∇ ξ û = ∇ ξ = ∇ξ x = .
∂t ∂t ∂t
Thus, we may write
∂
Ft
Ft +ε =
Ft + ε + o(ε) =
Ft + ε∇u
Ft + o(ε) = (I + ε∇u) Ft + o(ε).
∂t
We now exploit the well-known result that for any matrix A,
det(I + εA) = 1 + ε tr A + o(ε),

where tr A = i Aii denotes the trace of the matrix A, to write
Jt +ε = det( Jt + o(ε) = (1 + ε div û)Jt + o(ε).

Ft +ε ) = (1 + ε tr ∇u)
We have used the identity tr ∇u = div û. Then, by applying the definition of material
derivative and exploiting the continuity of the Lagrangian mapping, we may write

DJt Jt +ε − Jt
= lim ◦ L−1 −1
t = (div û Jt ) ◦ Lt = div u Jt .
Dt ε→0 ε
E XAMPLE 7.3. For the movement law given by Example 7.1, we have
t
e 0
Jt = det = et
0 1
and Jt = et as well. We may verify directly relation (7.10) since
d D
Jt div u = et (1 + 0) = et = Jt = (by relation (7.3)) = Jt .
dt Dt
We have now the following fundamental result.
T HEOREM 7.2 (Reynolds transport theorem). Let V0 ⊂ Ω0 , and Vt ⊂ Ωt be its image

under the mapping Lt . Let f : I × Ωt → R be a continuously differentiable function
with respect to both variables x and t. Then,

d Df ∂f
f= + f div u = + div(f u) . (7.11)
dt Vt Vt Dt Vt ∂t
P ROOF. Thanks to Theorem 7.1 and relations (7.10) and (7.3), we have

d d ˆ ∂ˆ
f (t, x) dx = f (t, ξ )Jt (ξ ) dξ = f (t, ξ )Jt (ξ ) dξ
dt Vt dt V0 V0 ∂t
ˆ
∂f ∂
= (t, ξ )Jt (ξ ) + fˆ(t, ξ ) Jt (ξ ) dξ .
V0 ∂t ∂t
We now use Theorem 7.1 and the definition of material derivative (7.3) to write

∂ fˆ Df
(t, ξ )Jt (ξ ) dξ = (t, x) dx.
V0 ∂t Vt Dt
Furthermore, we exploit again the definition of material derivative (7.3) in order to

rewrite relation (7.10) in the following equivalent form:
∂
Jt (ξ ) = Jt (ξ ) div u t, x(t, ξ) .
∂t
S ECTION 8 25
Consequently,

d Df
f (t, x) dx = (t, x) dx + fˆ(t, ξ )Jt x(t, ξ ) div u t, x(t, ξ ) dξ
dt Vt Vt Dt V0

Df
= (t, x) dx + f (t, x) div u(t, x) dx
Vt Dt Vt

Df
= (t, x) + f (t, x) div u(t, x) dx.
Vt Dt
The second equality in (7.11) is a consequence of (7.4).
Relation (7.11) is given the name of Reynolds transport formula, or simply transport
formula (sometimes the name convection formula is used as well).
By the application of the divergence theorem the previous expression becomes

d ∂f
f= + f u · n.
dt Vt Vt ∂t ∂Vt
8. The derivation of the basic equations of fluid mechanics
In the sequel, the symbol Vt will always be used to indicate a material volume at time t,
i.e., Vt is the image under the Lagrangian mapping of a subdomain V0 ⊂ Ω0 , i.e., Vt =
Lt (V0 ) (as already done in Theorem 7.2).
8.1. Continuity equation or mass conservation
We assume that there exists a strictly positive, measurable function ρ : I × Ωt → R,

called density such that on each Vt ⊂ Ωt ,

ρ = m(Vt ),
Vt
where m(Vt ) is the mass of the material contained in Vt . The density ρ has dimensions
[ρ] = kg/m3 .
A fundamental principle of classical mechanics, called principle of mass conserva-
tion, states that mass is neither created nor destroyed during the motion. This principle
translates into the following mathematical statement.
Given any material volume Vt ⊂ Ωt the following equality holds:

d
ρ = 0.
dt Vt
We can apply the transport theorem, obtaining

Dρ
+ ρ div u = 0. (8.1)
Vt Dt
By assuming that the terms under the integral are continuous, the arbitrariness of Vt
allows us to write the continuity equation in differential form
∂ρ
+ div ρu = 0.
∂t
In these cases for which we can make the assumption that ρ is constant (like for blood
flow), we obtain
div u = 0. (8.2)
Relation (8.2), which has been derived from the continuity equation in the case of a con-
stant density fluid (sometimes also called incompressible fluid), is indeed a kinematic
constraint. Thanks to (7.10), relation (8.2) is equivalent to
D
Jt = 0, (8.3)
Dt
which is the incompressibility constraint. A flow which satisfies the incompressibility
constraint is called incompressible. By the continuity equation, we derive the following
implication:
constant density fluid ⇒ incompressible flow,

whereas the converse is not true in general.
By employing the transport formula (7.11) with f = 1, we may note that the incom-
pressibility constraint is equivalent to

d
dx = 0 ∀Vt ⊂ Ωt ,
dt Vt
which means that the only possible motions of an incompressible flow are those which
preserve the fluid volume.
8.2. The momentum equation
Another important principle allows the derivation of an additional set of differential

equations, that is the principle of conservation of momentum. It is an extension of the
famous Newton law, “force = mass × acceleration”, to a continuous medium.
R EMARK 8.1. In the dimension unit specifications we will use the symbol Ne to indi-
cate the Newtons (the dimension units of a force), Ne = kg m/s2 , instead of the more
standard symbol N, since we have used the latter to indicate the number of space di-
mensions.
Three different types of forces may be acting on the material inside Ωt :

• Body forces. These forces are proportional to the mass. They are normally repre-
sented by introducing a vector field f b : I × Ωt → R3 , called specific body force,
whose dimension unit, [f b ] = Ne /kg = m/s2 , is that of an acceleration. The body
S ECTION 8 27
force acting on a volume Vt is given by

ρf b ,
Vt
whose dimension unit is clearly Ne. An example is the gravity force, given by
f b = −ge3 , where e3 represents the vertical direction and g the gravitational ac-
celeration.
• Applied surface forces. They represent that part of the forces which are imposed
on the medium through its surface. We will assume that they may be represented
through a vector field te : I × Γtn → R3 , called applied stresses, defined on a
measurable subset of the domain boundary Γtn ⊂ ∂Ωt and with dimension unit
[te ] = Ne /m2 . The resultant force acting through the surface is then given by

te .
Γtn
An example of a surface stress is that caused by the friction of the air flowing over
the surface of a lake.
• Internal “continuity” forces. These are the forces that the continuum media par-
ticles exert on each other and are responsible for maintaining material continuity
during the movement. To model these forces let us recall the following principle,
due to Cauchy.
8.2.1. The Cauchy principle

There exists a vector field t, called Cauchy stress,
t : I × Ωt × S1 → R3
with

S1 = n ∈ R3 : |n| = 1
such that its integral on the surface of any material domain Vt ⊂ Ωt , given by

t(t, x, n) dσ (8.4)
∂Vt
is equivalent to the resultant of the material continuity forces acting on Vt . In (8.4),
n indicates the outward normal of ∂Vt .
Furthermore, we have that
t = te on ∂Vt ∩ Γtn .
This principle is of fundamental importance because it states that the only dependence
of the internal forces on the geometry of ∂Vt is through n.
We may now state the following principle of conservation of linear momentum.
For any t ∈ I , on any sub-domain Vt ⊂ Ωt completely contained in Ωt , the following
relation holds:

d
ρ(t, x)u(t, x) dx = ρ(t, x)f b (t, x) dx + t(t, x, n) dσ, (8.5)
dt Vt Vt ∂Vt
where all terms dimension unit is Ne. Relation (8.5) expresses the property that the
variation of the linear momentum of Vt (represented by the integral at the left-hand
side) is balanced by the resultant of the internal and body forces.
With some further assumptions on the regularity of the Cauchy stresses, we are now
able to relate the internal continuity forces to a tensor field, as follows.
T HEOREM 8.1 (Cauchy stress tensor theorem). Let us assume that ∀t ∈ I , the body
forces f b , the density ρ and (D/Dt )u are all bounded functions on Ωt and that the
Cauchy stress vector field t is continuously differentiable with respect to the variable
x for each n ∈ S1 , and continuous with respect to n. Then, there exists a continuously
differentiable symmetric1 tensor field, called Cauchy stress tensor,
T : I × Ωt → R3×3 , [T] = Ne /m2 ,
such that
t(t, x, n) = T(t, x) · n, ∀t ∈ I, ∀x ∈ Ωt , ∀n ∈ S1 .
The proof is omitted. The interested reader may refer to A RIS [1962], S ERRIN [1959].
Therefore, under the hypotheses of the Cauchy theorem, we have
T · n = te , on ∂Vt ∩ Γtn , (8.6)
and that the resultant of the internal forces on Vt is expressed by

T · n, (8.7)
∂Vt
and we may rewrite the principle of conservation of linear momentum (8.5) as follows.
For all t ∈ I , on any sub-domain Vt ⊂ Ωt completely contained in Ωt , the following
relation holds:

d
ρu = ρf +
b
T · n. (8.8)
dt Vt Vt ∂Vt
Since ρ is constant and div u = 0, by invoking the transport formula (7.11), we obtain

d D Du
ρu = (ρu) + ρu div u = ρ .
dt Vt Vt Dt Vt Dt
By using the divergence theorem and assuming that div T is integrable, relation (8.8)
becomes

Du
ρ − div T − ρf b = 0.
Vt Dt
Thanks to the arbitrariness of Vt and under the hypothesis that the terms under the
integrals are continuous in space, we derive the following differential equation:
Du
ρ − div T = ρf b in Ωt . (8.9)
Dt
1 The symmetry of the Cauchy tensor may indeed be derived from the conservation of angular momentum.
S ECTION 8 29
R EMARK 8.2. In deriving (8.9), we have assumed that Vt is completely contained

into Ωt . We may however extend the derivation to the case where Vt has a part of
boundary in common with Γtn . In that case, we should use in place of (8.8) the follow-
ing:

d
ρu = T·n + te + ρf b . (8.10)
dt Vt ∂Vt \Γtn ∂Vt ∩Γtn Vt
Even now we would re-obtain (8.9) in view of property (8.6) of the Cauchy stress tensor,
which should now be regarded as boundary condition.
We may note that Du/Dt is indeed the fluid acceleration. Referring to relation (7.5),
it may be written as
Du ∂u
= + (u·∇)u,
Dt ∂t
where (u·∇)u is a vector whose components are

3
∂ui
(u·∇)u i = uj , i = 1, . . . , 3.
∂xj
j =1
For ease of notation, from now on we will omit the subscript b to indicate the body
force density applied to the fluid, which will be indicated just as f.
Relation (8.9) may finally be written as
∂u
ρ + ρ(u·∇)u − div T = ρf. (8.11)
∂t
Component-wise,
∂ui ∂ui ∂Tij

3 3
ρ +ρ uj − = ρfib , i = 1, . . . , 3.
∂t ∂xj ∂xj
j =1 j =1
The non-linear term ρ(u·∇)u is called the convective term.
R EMARK 8.3. We note the convective term may be written in the so-called divergence
form div(u ⊗ u), where

3
∂
(div u ⊗ u)i = (ui uj ), i = 1, . . . , 3.
∂xj
j =1
Indeed, thanks to the incompressibility of the fluid,
(u·∇)u = (u·∇)u + u div u = div(u ⊗ u).

The momentum equation in divergence form is then
∂u
ρ + ρ div(u ⊗ u − T) = ρf. (8.12)
∂t
8.3. The constitutive law
In order to close the system of Eqs. (8.2) and (8.11) just derived, we need to link the
Cauchy stress tensor to the kinematic quantities, and in particular, the velocity field.
Such a relation, called constitutive law, provides a characterization of the mechanical
behavior of the particular fluid under consideration.
The branch of science which studies the behavior of a moving fluid and in particular
the relation between stresses and kinematic quantities is called rheology. We have al-
ready anticipated in the introduction that blood rheology could be complex, particularly
in vessels with small size.
Here, we will assume for the fluid a Newtonian behavior (an approximation valid for
many fluids and also for blood flow in large vessels, which is the case in our presenta-
tion). In a Newtonian incompressible fluid, the Cauchy stress tensor may be written as
a linear function of the velocity derivatives (S ERRIN [1959]), according to

T = −P I + µ ∇u + ∇uT , (8.13)
where P is a scalar function called pressure, I is the identity matrix, µ is the dynamic
viscosity of the fluid and is a positive quantity. The tensor

(∇u + ∇uT ) 1 ∂ui ∂uj
D(u) = , Dij = + , i = 1, . . . , 3, j = 1, . . . , 3,
2 2 ∂xj ∂xi
is called the strain rate tensor. Then,
T = −P I + 2µD(u).
The term 2µD(u) in the definition of the Cauchy stress tensor is often referred to
as viscous stress component of the stress tensor. We have that [P ] = Ne /m2 and
[µ] = kg/m s. The viscosity may vary with respect to time and space. For example, it
may depend on the fluid temperature. The assumption of Newtonian fluid, however, im-
plies that µ is independent from kinematic quantities. Simple models for non-Newtonian
fluids, often used for blood flow simulations, express the viscosity as function of the
strain rate, that is µ = µ(D(u)). The treatment of such cases is rather complex and will
not be considered here, the interested reader may consult, for instance, R AJAGOPAL
[1993], C OKELET [1987].
We now recall that, if P is a scalar and Σ a vector field, then
div(P Σ ) = ∇P Σ + P div Σ,
and, therefore,
div(P I) = ∇P I + P div I = ∇P .
The momentum equation may then be written as
∂u
ρ + ρ(u·∇)u + ∇P − 2 div µD(u) = ρf.
∂t
S ECTION 9 31
Since ρ is constant, it is sometimes convenient to introduce the kinematic viscosity

ν = µ/ρ, with [ν] = m2 /s, and to write
∂u
+ (u·∇)u + ∇p − 2 div νD(u) = f, (8.14)
∂t
where p = P /ρ is a scaled pressure (with [p] = m2 /s2 ).
R EMARK 8.4. Under the additional hypothesis that ν is constant, the momentum equa-
tion may be further elaborated by considering that
div ∇u = u,
div ∇uT = ∇(div u) = (by relation (8.2)) = 0.
Consequently, the momentum equation for an incompressible Newtonian fluid with
constant viscosity may be written in the alternative form
∂u
+ (u·∇)u + ∇p − ν u = f. (8.15)
∂t
However, for reasons that will appear clear later on (and that have to see with the dif-
ferent natural boundary conditions associated with the two formulations), we prefer to
use the Navier–Stokes equations in the form (8.14), even when considering a constant
viscosity.
9. The Navier–Stokes equations
The set of differential equations formed by the continuity equation and the momen-
tum equations in the form derived in the previous section provides the Navier–Stokes
equations for incompressible fluids.
They are, in particular, valid on any fixed spatial domain Ω which is for all times
of interest inside the portion of space filled by the fluid, i.e., Ω ⊂ Ωt . Indeed, in most
cases, as with the flow around a car or an aeroplane, the flow motion is studied in a fixed
domain Ω (usually called computational domain) embodying the region of interest. We
will see in Section 18 that this is not possible anymore when considering the fluid–
structure interaction problem arising when blood is flowing in a large artery.
Yet, before addressing this more complex situation, we will analyze the Navier–
Stokes equations in a fixed domain, that is, we will consider, for any t ∈ I , the system
of equations
∂u
+ (u·∇)u + ∇p − 2 div νD(u) = f, in Ω,
∂t
div u = 0, in Ω. (9.1)
Furthermore, we need to prescribe the initial status of the fluid velocity, for instance
u(t = t0 , x) = u0 (x), x ∈ Ω. (9.2)

The principal unknowns are the velocity u and the “scaled” pressure p = P /ρ.
Let us take a practical case-study, namely the blood flow in an artery, for example the
carotid (ref. Fig. 2.3), which we will here consider rigid. We proceed by identifying the
area of interest, which may be the carotid sinus, and a domain Ω which will contain that
area and which extends into the vessels up to a certain distance. For obvious practical
reasons we will need to “truncate” the domain at certain sections. Inside such domain,
the Navier–Stokes equations are valid, yet in order to solve them we need to provide
appropriate boundary conditions.
9.1. Boundary conditions for the Navier–Stokes equations
The Navier–Stokes equations must be supplemented by proper boundary conditions that

allow the determination of the velocity field up to the boundary of the computational do-
main Ω. The more classical boundary conditions which are mathematically compatible
with the Navier–Stokes equations are:
(1) Applied stresses (or Neumann boundary condition). We have already faced this
condition when discussing the Cauchy principle. With the current definition for
the Cauchy stresses it becomes
T · n = −P n + 2µD(u) · n = te on Γ n ⊂ ∂Ω, (9.3)

where Γ n is a measurable subset (possibly empty) of the whole boundary ∂Ω.
(2) Prescribed velocity (or Dirichlet boundary condition). A given velocity field is
imposed on Γ d , a measurable subset of ∂Ω (which may be empty). This means
that a vector field
g : I × Γ d → R3
is prescribed and we impose that
u=g on Γ d .
Since div u = 0 in Ω, it must be noted that if Γ d = ∂Ω then at any time g must
satisfy the following compatibility condition:

g · n = 0. (9.4)
∂Ω
Clearly, for a proper boundary conditions specification we must have Γ n ∪ Γ d = ∂Ω.

The conditions to apply are normally driven by physical considerations. For instance,
for a viscous fluid (µ > 0) like the one we are considering here, physical considerations
lead to impose the homogeneous Dirichlet condition u = 0 at a solid fixed boundary.
When dealing with an “artificial boundary”, that is a boundary which truncates the space
occupied by the fluid (for computational reasons) the choice of appropriate conditions is
often more delicate and should in any case guarantee the well-posedness of the resulting
differential problem.
For example, for the flow field inside a 2D model for the carotid artery such as the
one shown in Fig. 9.1, we could impose a Dirichlet boundary condition on Γ in , by
prescribing a velocity field g.
S ECTION 9 33
F IG . 9.1. A possible boundary subdivision for the flow in a carotid bifurcation.
On the “wall” boundary Γ w , which is in this case assumed to be fixed, we will im-
pose homogeneous Dirichlet conditions, that is u = 0 on Γ w . When we will consider
the coupled problem between fluid and vessel wall, Γ w will be moving, hence the ho-
mogeneous Dirichlet condition will be replaced by u = w, where w is the wall velocity.
At the exit Γ out , we could, for instance, impose homogeneous Neumann conditions,
i.e., relation (9.3) with te = 0. For the case illustrated in Fig. 9.1 and with that choice of
coordinate basis, it becomes (derivation left as exercise)

∂u1 ∂u2 ∂u2
µ + = 0, −P + 2µ = 0.
∂x2 ∂x1 ∂x2
R EMARK 9.1. We anticipate the fact (without providing the proof) that this choice of
boundary conditions, with the hypothesis that at Γ out the velocity satisfies everywhere
the condition u · n > 0, is sufficient to guarantee that the solution of the Navier–Stokes
problem exists and is continuously dependent from the data (initial solution, boundary
conditions, forcing terms), provided that the initial data and forcing term are sufficiently
small.
Unfortunately, the homogeneous Neumann condition, which indeed would simulate

a discharge into the open air, is rather unphysical for the case of a human vessel. As
a matter of fact, it neglects completely the presence of the remaining part of the cir-
culatory system. The difficulty in devising proper boundary conditions for this specific
problem was already mentioned in Section 2 of these notes. The matter is still open
and is the subject of active research. A possibility is provided by coupling the Navier–
Stokes equations on the section of the arterial tree of interest with reduced models,
like the one that will be presented in Section 20, which are able to represent, though in a
simplified way, the presence of the remaining part of the circulatory system. Techniques
of this type has been used and analysed in F ORMAGGIA , N OBILE , Q UARTERONI and
V ENEZIANI [1999], F ORMAGGIA , G ERBEAU , N OBILE and Q UARTERONI [2001].
C HAPTER III
10. The incompressible Navier–Stokes equations and their approximation
In this section we introduce the weak formulation of the Navier–Stokes equations for
constant density (incompressible) fluids. Then, we address basic issues concerning the
approximation of these equations in the context of the finite element method.
10.1. Some functional spaces
For the following discussion we need to introduce some Sobolev spaces for vector func-
tions. We assume that the reader is already acquainted with the main definitions and re-
sults on Sobolev spaces in one dimension. A simple introduction is provided in R EDDY
[1998]. For a deeper insight see, for instance, B REZIS [1983].
We will indicate with Lp (Ω) (1 p ∞) the space of vector functions f : Ω → RN
(with N = 2 or 3) whose components belong to Lp (Ω). Its norm is
N 1/p
p
fLp (Ω) = fi Lp (Ω) , 1 p < ∞,
i=1
and

fL∞ (Ω) = inf C ∈ R: |fi | C, i = 1, . . . , N, a.e. in Ω ,
where “a.e.” stands for “almost everywhere”. We will use the same notation for tensor
fields, i.e., we will also indicate with Lp (Ω) the space of tensor fields T : Ω → RN×N
whose components belongs to Lp (Ω). In this case
N N 1/p
p
TLp (Ω) = Tij Lp (Ω) , 1 p < ∞.
i=1 j =1
Analogously, a vector (or a tensor) function f belongs to Hm (Ω) if all its components
belong to H m (Ω), and we have
N 1/2

fHm (Ω) = fi H m (Ω)
2
,
i=1
while its semi-norm is
N 1/2

|f|Hm (Ω) = |fi |H m (Ω)
2
.
i=1
35
36 A. Quarteroni and L. Formaggia C HAPTER III
It is understood that, when m = 0,
H0 (Ω) ≡ L2 (Ω).
When equipped with the following scalar product:

N
(f, g)Hm (Ω) = (fi , gi )H m (Ω) , f, g ∈ Hm (Ω),
i=1
the space Hm (Ω)is a Hilbert space.

To ease notation, we will often use the following short-hand notation for the L2 scalar
products:
(v, w) ≡ (v, w)L2 (Ω) , (p, q) ≡ (p, q)L2 (Ω) .
We note that the L2 scalar product of two tensor fields T and G belonging to L2 (Ω)
is defined as
N
N
(T, G) ≡ (T, G)L2 (Ω) = T:G= Tij Gij .
Ω i=1 j =1 Ω
For our purposes we will usually have m = 1. In that case we have the equality
f2H1 (Ω) = f2L2 (Ω) + ∇ f2L2 (Ω) .
We often utilise the space H10 (Ω) defined as

H10 (Ω) = v ∈ H1 (Ω): v|∂Ω = 0 .
We will consider bounded domains Ω with regular (i.e., Lipschitz continuous) bound-
ary ∂Ω, so that both the Sobolev embedding theorems in RN and the Green integration
formula hold. Some important results are here recalled, without providing the demon-
stration, which may be found in L IONS and M AGENES [1968] or B REZIS [1983].
T HEOREM 10.1 (Sobolev embeddings (simplified form)). Let Ω be a bounded domain

of RN with Lipschitz continuous boundary. The following properties hold:

 N 2N

 If 0 s < 2 , H (Ω)
→ L (Ω), p = N − 2s ,
s p



N
If s = , Hs (Ω)
→ Lq (Ω), 2 q < ∞,

 2


 If s > N , Hs (Ω)
→ C 0 (Ω)N ,

2
where A
→ B means that A is included in B with continuous embedding.
T HEOREM 10.2 (Green integration formula). Let Ω be a bounded domain of RN with

Lipschitz continuous boundary and let n denote the unit outer normal along ∂Ω. Let
S ECTION 11 37
u, v ∈ H 1 (Ω), then the integral

uv ni
∂Ω
exists and is finite for each component ni of n. In addition we have

∂u ∂v
v=− u + uvni , i = 1, . . . , N.
Ω ∂xi Ω ∂xi ∂Ω
L EMMA 10.1 (Poincaré inequality – multidimensional case). Let f : RN → R be a

function of H1 (Ω), with f = 0 on Γ ⊂ ∂Ω of strictly positive measure. Then there
exists a positive constant CP (depending only on the domain Ω and on Γ ), such that
f L2 (Ω) CP ∇f L2 (Ω) . (10.1)
L EMMA 10.2. Let Ω be a bounded and connected subset of RN , where N = 2 or 3.

Furthermore, let us assume that the velocity field u ∈ H1 (Ω) vanishes on Γ ⊂ ∂Ω of
strictly positive measure. Then, there exists a constant CK > 0 so that the following
inequality holds:

D(u) : D(u) CK ∇u2L2 (Ω) . (10.2)
Ω
This theorem is a consequence of the Korn inequality, whose precise statement may
be found, for instance, in C IARLET [1988], D UVAUT and L IONS [1976].
L EMMA 10.3 (Gronwall lemma). Let f be a non-negative function which is integrable

in I = (t0 , t1 ) and g and φ be two continuous functions in I , with g non-decreasing. If
t
φ(t) g(t) + f (τ )φ(τ ) dτ ∀t ∈ I, (10.3)
t0
then
t
φ(t) g(t) exp f (τ ) dτ ∀t ∈ I. (10.4)
t0
11. Weak form of Navier–Stokes equations
The incompressible Navier–Stokes equations read

∂u
+ (u·∇)u + ∇p − 2 div νD(u) = f, in Ω, t ∈ I, (11.1a)
∂t
div u = 0, in Ω, t ∈ I, (11.1b)
u = u0 , in Ω, t = t0 . (11.1c)
We assume that ν is a bounded strictly positive function, precisely we assume that there
exist two constants ν0 > 0 and ν1 > 0 such that ∀t ∈ I ,
ν0 ν ν1 almost everywhere in Ω.
We consider the case in which the system of differential equations (11.1) is equipped
with the following boundary conditions:
u = g on Γ d , t ∈ I, (11.2a)
−pn + 2νD(u) · n = h on Γ , t ∈ I,
n
(11.2b)
We have indicated with Γ d and Γ n the portions of ∂Ω where Dirichlet and Neumann
boundary conditions are applied, respectively. We must have Γ d ∪ Γ n = ∂Ω.
R EMARK 11.1. If Γ d = ∂Ω we call the problem formed by (11.1) and (11.2) a Dirich-
let problem. We will instead use the term Neumann problem when Γ n = ∂Ω. The con-
ditions g = 0 and h = 0 are called homogeneous boundary conditions.
In the case of a Dirichlet problem, the boundary datum has to satisfy the following
compatibility relation for all t ∈ I :

g · n = 0.
∂Ω
R EMARK 11.2. For the problem at hand, we normally have f = 0, since the only ex-
ternal force which one may eventually consider in blood flow is the gravity force. Even
in this case, we may still adopt the Navier–Stokes equations with f = 0 by replacing p
with p∗ (t, x) = p(t, x) + gz(x)ez , where g is the gravity acceleration, ez the unit vector
defining the vertical direction (upwardly oriented) and z(x) the (known) quota of point
x with respect to a reference horizontal plane. Yet, for the sake of completeness, many
of the derivations of this as well as the following sections refer to the general case f = 0.
The weak form of the Navier–Stokes equations is (formally) obtained by taking the
scalar product of the momentum equations with a vector function v belonging to a
functional space V (called test function space), which will be better specified later on,
integrating over Ω and applying the Green integration formula. We operate similarly on
the continuity equation, by multiplying it by a function q ∈ Q and integrating. Also the
space Q will be specified at a later stage.
We formally obtain

∂u
, v + (u·∇)u, v + 2 νD(u) : D(v) − (p, div v)
∂t Ω

= (f, v) + v · 2νD(u) · n − pn ,
∂Ω
(div u, q) = 0.
We have exploited the identity

νD(u) : ∇v = νD(u) : D(v),
Ω Ω
which derives from the symmetry of the tensor D(u).
S ECTION 11 39
The boundary term may now be split into two parts,

v · 2νD(u) · n − pn = v · 2νD(u) · n − pn + v · h.
∂Ω Γd Γn
We note that the contribution from the Neumann boundary is now a given data, while
contribution from the Dirichlet boundary can be eliminated by appropriately choosing
the test space V.
By inspection, we may recognise that all terms make sense if we choose as test func-
tion spaces

V = v ∈ H1 (Ω): v|Γ d = 0 ,

Q = q ∈ L2 (Ω): with q = 0 if Γ d = ∂Ω ,
Ω
and if we seek, at each time t, the velocity in

Vg = u ∈ H1 (Ω): u|Γ d = g
and the pressure in Q.
R EMARK 11.3. The request that Q is formed by functions with zero mean on Ω when
we treat a Dirichlet problem derives from the fact that in such a case the pressure is
determined only up to a constant, as it appears in the equations only through its gradient.
To compute a unique value for the pressure it is then necessary to fix the constant. This
is obtained by the zero-mean constraint.
Finally, the weak form of the Navier–Stokes problem (11.1) and (11.2), reads:
Find, ∀t ∈ I , u(t) ∈ Vg and p(t) ∈ Q such that

 ∂u
, v + a(u, v) + c(u, u, v) + b(v, p) = (f, v) + Γ n v · h, ∀v ∈ V,
∂t (11.3)

b(u, q) = 0, ∀q ∈ Q,
where

a(u, v) = 2 νD(u) : D(v), (11.4)
Ω
c(w, u, v) = (w·∇)u · v, (11.5)
Ω
b(v, p) = − p div v. (11.6)
Ω
11.1. The homogeneous Dirichlet problem
In this section we will focus on the homogeneous Dirichlet problem, that is the case
when Γ d = ∂Ω and g = 0 in (11.2a). Therefore,

V = H10 (Ω), Q = L20 (Ω) = q ∈ L2 (Ω), q =0 (11.7)
Ω
and the weak form reads:

Find, ∀t ∈ I , u(t) ∈ V and p(t) ∈ Q such that

 ∂u
, v + a(u, v) + c(u, u, v) + b(v, p) = (f, v), ∀v ∈ V,
∂t (11.8)

b(u, q) = 0, ∀q ∈ Q.
L EMMA 11.1. The forms a : V × V → R, c : V × V × V → R and b : V × Q → R are

continuous with respect to their arguments. In addition, a(·, ·) is coercive, i.e., ∃α > 0
such that
a(v, v) αv2H1 (Ω) , ∀v ∈ V.
P ROOF. The continuity of the bilinear forms a and b is an immediate consequence of

the Cauchy–Schwarz inequality. Indeed, ∀u, v ∈ V and ∀q ∈ Q,

a(u, v) ν1 |u| 1 |v| 1
H (Ω) H (Ω) ν1 uH1 (Ω) vH1 (Ω) ,

b(u, p) div u 2 p 2 u 1 p 2 .
L (Ω) L (Ω) H (Ω) L (Ω)
For the tri-linear form c we first have to note that thanks to the Sobolev embedding
theorem H1 (Ω)
→ L6 (Ω) (as N = 2, 3) and consequently H1 (Ω)
→ L4 (Ω). Then,
wu ∈ L2 (Ω), and considering the expression of c(·, ·, ·) component-wise, we have

∂uk ∂uk ∂uk
wi
vk wi vk L2 (Ω)
wi L4 (Ω) vk L4 (Ω) .
Ω ∂xi ∂xi L2 (Ω) ∂xi L2 (Ω)
Then

∂uk ∂uk
wi
vk Cwi H 1 (Ω) vk H 1 (Ω)
Ω ∂xi ∂xi L2 (Ω)
Cwi H 1 (Ω) |uk |H 1 (Ω) vk H 1 (Ω)
Cwi H 1 (Ω) uk H 1 (Ω) vk H 1 (Ω) , (11.9)
where C is a positive constant.
It follows that, ∀u, v, v ∈ V,
c(w, u, v) C1 wH1 (Ω) uH1 (Ω) vH1 (Ω) ,
by which the continuity of the tri-linear form is proved (C1 is a positive constant).
The coercivity of the linear form a derives from inequalities (10.1) and (10.2), since

a(v, v) 2ν0 D(v) : D(v) 2ν0 CK |v|2H1 (Ω) αv2H1 (Ω) , ∀v ∈ V, (11.10)
Ω
with α = (2ν0 CK )/(CP2 + 1), being CP and CK the constants in (10.1) and (10.2),
respectively.
We now introduce the space

Vdiv = {v ∈ V : div v = 0 a.e. in Ω}.
S ECTION 12 41
T HEOREM 11.1. If u is a solution of the weak formulation (11.8), then u(t) ∈ Vdiv for
all t ∈ I and it satisfies

∂u
, v + a(u, v) + c(u, u, v) = (f, v), ∀v ∈ Vdiv , t ∈ I. (11.11)
∂t
Conversely, if, ∀t ∈ I , u(t) ∈ Vdiv is a solution of (11.11) and ∂u/∂t ∈ L2 (Ω), then
there exists a unique p ∈ Q such that (u, p) satisfies (11.8).
P ROOF. The first part of the proof is trivial. If u satisfies (11.8) then it belongs to Vdiv
and it satisfies (11.11), since Vdiv ⊂ V.
The demonstration of the inverse implication requires first to state the following re-
sult.
L EMMA 11.2. Let Ω be a domain of RN and let L ∈ V . Then L(v) = 0, ∀v ∈ Vdiv if

and only if there exists a function p ∈ L2 (Ω) such that
L(v) = (p, div v), ∀v ∈ V.
For the proof see Lemma 2.1 of G IRAULT and R AVIART [1986].
The application L defined as

∂u
L(v) = , v + a(u, v) + c(u, u, v) − (f, v), ∀v ∈ V,
∂t
belongs to V , being a linear continuous functional on V. We can therefore apply
Lemma 11.2 and obtain the desired result.
12. An energy inequality for the Navier–Stokes equations
We now prove an energy inequality for problem (11.8), by which we may assess a
continuous dependence of the solution from the given data.
T HEOREM 12.1 (Energy inequalities). Let u(t) ∈ Vdiv be a solution of (11.8), ∀t ∈ I .

Then the following inequalities hold:
t t
2 2
u(t)2 2 + C1
∇u(τ ) dτ u 2
0 L2 (Ω) + f(τ ) dτ et ,
L (Ω) L2 (Ω) L2 (Ω)
0 0
where C1 = 4ν0 CK , and
t
2 CP t
u(t)2 2 + C 2
∇u(τ ) 2 dτ u0 2
2 + f(τ )2 2 dτ,
L (Ω) L (Ω) L (Ω) C2 0 L (Ω)
0
where C2 = 2ν0 CK . Here, CK and CP are the constants in the Poincaré inequality
(10.1) and in (10.2), respectively.
We first prove the following result.

L EMMA 12.1. If u is a solution of (11.8) then c(u, u, u) = 0.
P ROOF. It follows from the Green formula and the fact that u|∂Ω = 0. Indeed,

1 2
c(u, u, u) = (u·∇)u · u = ∇ |u| · u
Ω 2
Ω

1 1
=− |u|2 div u + |u|2 u · n.
2 Ω 2 ∂Ω
Now, the last integral is zero since u = 0 on ∂Ω. Moreover, for the same reason

div u = u · n = 0.
Ω ∂Ω
Then, if we set

c= |u|2 ,
Ω
we have

|u| div u =
2
|u| div u − c
2
div u = |u|2 − c div u
Ω Ω Ω Ω
2
= b u, |u| − c = 0,
where the last equality is obtained since (|u|2 − c) ∈ Q and b(u, q) = 0, ∀q ∈ Q.
We now give the demonstration of Theorem 12.1.
P ROOF. For all fixed t, take v = u(t) in the momentum equation of (11.11). We have
1 d
u2L2 (Ω) + c(u, u, u) + b(u, p) + a(u, u) = (f, u). (12.1)
2 dt
Then,
1 d
u2L2 (Ω) + a(u, u) = (f, u).
2 dt
Now, thanks to (10.2),

a(u, u) = 2 νD(u) : D(u) 2ν0 CK ∇u2L2 (Ω) ,
Ω
then
d 1
u2L2 (Ω) + 4ν0 CK ∇u2L2 (Ω) 2(f, u) f2L2 (Ω) + 2εu2L2 (Ω) , (12.2)
dt 2ε
for any ε > 0. By choosing ε = 1/2 and integrating between t0 and t, we have
t

u(t)2 2 + 4ν0 CK
∇u(τ )2 2 dτ
L (Ω) L (Ω)
t0
t
t
f(τ )2 2 dτ + u(τ )2 2 dτ + u0 2L2 (Ω) .
L (Ω) L (Ω)
t0 t0
S ECTION 13 43
We apply Gronwall lemma (Lemma 10.3) by identifying

t

u(t)2 2 + 4ν C
0 K
∇u(τ )2 2 dτ
L (Ω) L (Ω)
t0
with φ(t), obtaining the first inequality.
By using instead the Poincaré inequality on the last term of (12.2), and by taking
ε = (ν0 CK )/CP2 , we obtain
d CP2
u2L2 (Ω) + 2ν0 CK ∇u2L2 (Ω) f2L2 (Ω) .
dt 2ν0 CK
By integrating between t0 and t, we obtain the second inequality of the theorem.
R EMARK 12.1. In the case where f = 0 we may derive the simpler estimate
t

u(t)2 2 + 4ν C
0 K
∇u(τ )2 2 dτ u0 2L2 (Ω) , ∀t t0 .
L (Ω) L (Ω)
0
13. The Stokes equations
The space discretisation of the Navier–Stokes equations give rise to a non-linear set
of ordinary differential equations because of the presence of the convective term. This
makes both the analysis and the numerical solution more difficult. In some cases, when
the fluid is highly viscous, the contribution of the non-linear convective term may be
neglected. The key parameter which allow us to make that decision is the Reynolds
number Re, which is an a-dimensional number defined as
|u|L
Re = ,
ν
where L represents a length-scale for the problem at hand and |u| the Euclidean norm
of the velocity. For the flow in a tube L is the tube diameter.
Contrary to other fluid dynamic situations, the high variation in time and space of
the velocity in the vascular system does not allow to select a single representative value
of the Reynolds number,2 nevertheless in the situations where Re 1 (for instance,
flow in smaller arteries or capillaries) we may say that the convective term is negligible
compared to the viscous contribution and may be discarded. We have then the Stokes
equations, which read (in the case of homogeneous Dirichlet conditions):
∂u
+ ∇p − 2 div νD(u) = f, in Ω, t ∈ I, (13.1a)
∂t
div u = 0, in Ω, t ∈ I, (13.1b)
u = 0, on ∂Ω, t ∈ I, (13.1c)
u = u0 , in Ω, t = t0 . (13.1d)
The corresponding weak form reads:
2 Another a-dimensional number which measures the relative importance of inertia versus viscous in oscil-
latory flow is the Womersley number (F UNG [1984]).
Find, ∀t ∈ I , u(t) ∈ V, p(t) ∈ Q, such that

∂u
, v + a(u, v) + b(v, p) = (f, v), ∀v ∈ V,
∂t
b(u, q) = 0, ∀q ∈ Q. (13.2)
In the case of a steady problem, that is when we consider ∂u/∂t = 0, the solution
(u, p) of the Stokes problem (13.2) is a saddle point for the functional
1
S(v, q) = a(v, v) + b(v, q) − (f, v), v ∈ V, q ∈ Q.
2
This means
S(u, p) = min max S(v, q).
v∈V q∈Q
In this respect, the pressure p may be considered as a Lagrange multiplier associated to

the incompressibility constraint.
R EMARK 13.1. In those cases where Re 1 (high Reynolds number flows) the flow
becomes unstable. High frequency fluctuations in the velocity and pressure field appear,
which might give rise to turbulence. This phenomenon is particularly complex and its
numerical simulation may be extremely difficult. To make the problem amenable to
numerical solution it is often necessary to adopt a turbulence model, which allows to
give a more or less accurate description of the effect of turbulence on the main flow
variables.
In normal physiological situations, the typical values of the Reynolds number reached
in the cardiovascular system do not allow the formation of full scale turbulence. Some
flow instabilities may occur only at the exit of the aortic valve and limited to the systolic
phase. Indeed, in this region the Reynolds number may reach the value of few thousands
only for the portion of the cardiac cycle corresponding the peak systolic velocity. There-
fore, there is no sufficient time for a full turbulent flow to develop.
The situation is different in some pathological circumstances, e.g., in the presence
of a stenotic artery. The increase of the velocity at the location of the vessel restriction
may induce turbulence to develop. This fact could explain the high increase in the noise
caused by the blood stream in this situation.
14. Numerical approximation of Navier–Stokes equations
In this section we give a very short account on possible numerical methods for the
solution of the Navier–Stokes equations. This subject is far from being simple, and
we will not make any attempt to be exhaustive. The interested reader can consult, for
instance, Q UARTERONI and VALLI [1994], Chapters 9, 10 and 13, and the classic books
on the subject by G IRAULT and R AVIART [1986] and T EMAM [1984].
Here, we will simply mention a few methods to advance the Navier–Stokes equations
from a given time-level to a new one and we will point out some of the mathematical
problems that have to be faced. For the sake of simplicity we will confine ourselves to
the homogeneous Dirichlet problem (11.8).
S ECTION 14 45
14.1. Time advancing by finite differences
The Navier–Stokes problem (9.1) (equivalently, its weak form (11.8)) can be advanced
in time by suitable finite difference schemes.
The simulation will cover the interval I = (0, T ) which we subdivide into sub-
intervals (time-steps) I k = (t k , t k+1 ) with k = 0, . . . , N and where t k+1 − t k = t is
constant. We have thus partitioned the space-time domain I × Ω into several time-slabs
I k × Ω. We assume that on each slab we know the solution at t = t k and that we wish to
find the solution at t = t k+1 . Clearly, for the first time slab the assumption is true since
at t = 0 the approximate solution is obtained from the initial data. If we treat the time
slabs in their natural order, as soon as the solution on the kth time slab has been found,
it is made available as initial condition for the computation on the next time slab. This
is a time-advancing procedure.
We will indicate by (uk , pk ) the approximate solution at time t k , that is
k k k k
u ,p ≈ u t ,p t .
A family of simple time-advancing schemes is obtained by using the Taylor expansion
formula to write

∂u k+1 u t k+1 − u(t k )
t = + O( t).
∂t t
Then, by making the first order approximation
∂u k+1 uk+1 − uk
t ≈ ,
∂t t
into (9.1), we may write the following time-stepping scheme to calculate uk+1 and
pk+1 :
uk+1 − uk
− 2 div νD uk+1 + (u∗ ·∇)u∗∗ + ∇pk+1 = f k+1 , in Ω, (14.1a)
t
div uk+1 = 0, in Ω, (14.1b)
uk+1
= 0, on ∂Ω. (14.1c)
Here, f k+1 stands for f(t k+1 ).
The value of u∗ and u∗∗ in the non-linear convective term may be taken, for instance,
as follows:
 k k
 u ·∇ u , fully explicit treatment,
∗ ∗∗  k k+1
u ·∇ u = u ·∇ u , semi-implicit treatment,

 k+1 k+1
u ·∇ u , fully implicit treatment.
In the case of the fully implicit treatment, Eqs. (14.1) give rise to a non-linear system.
The semi-implicit and fully explicit treatments, instead, perform a linearisation of the
convective term, thus eliminating the non-linearity.
Let us consider the scheme resulting from the fully explicit treatment of the convective
term. Problem (14.1) is then rewritten as
1 k+1 1 k k k
u − 2 div νD uk+1 + ∇pk+1 = f k+1 + u − u ·∇ u in Ω,
t t
(14.2a)
div uk+1 = 0, in Ω, (14.2b)
uk+1 = 0, on ∂Ω. (14.2c)
We will now denote uk+1 and pk+1 by w and π , respectively, and by q and a0 the
quantities
1 k k k 1
q = f k+1 + u − u ·∇ u , a0 = . (14.3)
t t
Problem (14.2) may be written in the form

a0 w − 2 div νD(w) + ∇π = q, in Ω, (14.4a)
div w = 0, in Ω, (14.4b)
w = 0, on ∂Ω, (14.4c)
which is called the generalised Stokes problem.
A characteristic treatment of the time derivative would also lead at each time step to
a generalised Stokes problem (see Section 14.3).
For its approximation, a Galerkin finite element procedure can be set up by consider-
ing two finite element spaces Vh for the velocity and Qh for the pressure, and seeking
wh ∈ Vh and πh ∈ Qh such that

ã(wh , vh ) + b(wh , πh ) = (q, vh ), ∀vh ∈ Vh ,
(14.5)
b(wh , qh ) = 0, ∀qh ∈ Qh ,
where ã(w, v) = a0 (w, v) + a(w, v).
The algebraic form of problem (14.5) is derived by denoting with
{ϕ i , i = 1, . . . , NVh }, {ψi , i = 1, . . . , NQh }

the bases of Vh and Qh , respectively. Here NVh = dim(Vh ) and NQh = dim(Qh ). Then,
by setting
NVh NQh

wh (x) = wi ϕ i (x), ph (x) = πi ψi (x), (14.6)
i=1 i=1
we obtain the following system from (14.5):

C DT W Fs
= , (14.7)
D 0 Π 0
where W, Π and Fs denote three vectors defined respectively as
(W)i = wi , (Π)i = πi , (Fs )i = (q, ϕ i ),

S ECTION 14 47
while C, K and D are matrices whose components are defined as

(C)ij = ã(ϕ j , ϕ i ), (D)ij = b(ϕ j , ψi ).
The global matrix

C DT
A= (14.8)
D 0
is a square matrix with dimension (NVh + NQh ) × (NVh + NQh ).
In the case of a finite element approximation, pi represents the pressure at the ith
mesh node. The interpretation of wi is made more complex by the fact that the velocity
is a vector function, while wi is a scalar. Let us assume that we are considering a three-
dimensional problem and let the basis for Vh be chosen by grouping the vector functions
ϕ i into 3 families, as follows:

ϕi 0 0
ϕi = 0 ,
1
ϕ i = ϕi ,
2
ϕi = 0 .
3
0 0 ϕi
Finally, let MVh = NVh /3. Then, we may rewrite the first expansion in (14.6) as
MVh

3
j j
wh (x) = wi ϕ i (x),
i=1 j =1
j
where wi here represents the j th component of w at the ith mesh node.
L EMMA 14.1. If ker D T = 0, then matrix A is non-singular.
P ROOF. We first prove the non-singularity of C. For any W ∈ RNVh , W = 0,

NVh NVh

W CW =
T
wi wj Cij = ã(w, w) > 0,
i=1 j =1
NV
where w = i=1h wi ϕ i . Consequently, C is positive-definite, and thus non-singular.
From (14.7) we have

W = C −1 Fs − D T Π , DW = 0.
Then we may formally compute the discrete pressure terms by

− DC −1 D T Π = −DC −1 Fs .
Proving that A is non-singular thus reduces to show that the matrix
S = DC −1 D T
is non-singular. If we take any q ∈ RNQh with |q| = 0 we have by hypothesis that
D T q = 0. Then
T
qT Sq = D T q C −1 D T q = 0,
since C −1 is symmetric positive definite. Thus, matrix S (which is clearly symmetric)

has all eigenvalues different from zero and, consequently, is non-singular. This con-
cludes the proof.
The scheme we have presented, with an explicit treatment of just the convective term,
is only one of the many possible ways of producing a time discretisation of the Navier–
Stokes equations. Another choice is to resort to a fully implicit scheme.
14.2. Fully implicit schemes
By employing in (14.1) a full implicit treatment of the convective part, we would obtain
a non-linear system of the following type:

E(W) D T W Fs
= , (14.9)
D 0 Π 0
where now the matrix E is a function of the unknown velocity,
NV
h
E(W) ij
= ã(ϕ i , ϕ j ) + c uk+1 , ϕ j , ϕ i = Cij + c(ϕ m , ϕ j , ϕ i )Wm .
m=1
A possible way to solve it is to resort to Newton’s method:

0
Given WΠ 0 , solve for l = 0, . . . ,
∂E
Wl · Wl + E Wl D T Wl+1 − Wl
∂W Π l+1 − Π l
D 0
l l
Fs E W DT W
= − , (14.10)
0 D 0 Πl
until a suitable convergence criterion is met.
The solution of a non-linear system is now reduced to a series of solutions of linear
systems. Going back to the Navier–Stokes equations, we may note that a full implicit
scheme would require to solve at each time step a series of linear systems of form
(14.10), that resembles the Stokes problem. The resulting numerical scheme is thus in
general very computationally intensive.
14.3. Semi-Lagrangian schemes
An alternative way to treat the non-linear term in the Navier–Stokes equations is ob-
tained by performing an operator splitting that separates the effect of the convective
term. The technique exploits the fact that the convective term is indeed the material
derivative of u,
∂u Du
+ u · ∇u = ,
∂t Dt
S ECTION 14 49
that is the derivative along the particle trajectories Tξ (also called characteristic lines)
defined in Section 7.
On each time-slab I k we then have that
t k+1
Du(τ, x)
dτ = u t k+1 , x − u t k+1 , x∗ ≈ uk+1 (x) − uk (x∗ ), (14.11)
tk Dτ
where x∗ is position at t = t k of the fluid particle located in x at t = t k+1 , i.e., x∗ = yx (1)
where yx (s) is the solution of

 dyx (s)
= −u t k+1 − s t, yx (s) ,
ds (14.12)

yx (0) = x.
The point x∗ is often denoted as the “foot” of the characteristic line yx .
This interpretation leads to the semi-Lagrangian schemes, so called because we treat
the convective operator in the Lagrangian frame. For instance, a backward Euler semi-
Lagrangian scheme will lead at each time step I k a generalised Stokes problem like
(14.4), where now
1 k ∗
q(x) = f k+1 (x) + u (x ),
t
that may then be treated by a Galerkin finite element procedure as described in Sec-
tion 14.1.
Clearly, system (14.12) has to be approximated as well. A first-order approximation
leads to
x∗ = x − uk (x) t.
This explicit treatment will eventually entail a stability condition which depends on the
fluid velocity. Higher-order schemes may be devised as well, see, for instance, B OUKIR ,
M ADAY, M ÉTIVET and R AZAFINDRAKOTO [1997].
The major drawback of semi-Lagrangian schemes is the computation of the approxi-
mation of uk (x∗ ). In a finite element context it requires to locate the mesh element where
the foot of the characteristic passing through each mesh point lies (or each quadrature
point if a quadrature rule is used to compute the space integrals). An efficient imple-
mentation calls for the use of special data structures. Furthermore, a proper treatment
is needed when x∗ falls outside the computational domain. In that case the boundary
conditions have to be properly taken into account.
14.4. Projection methods
We now follow another route for the solution of the incompressible Navier–Stokes equa-
tions which does not lead to a Stokes problem but to a series of simpler systems of partial
differential equations. We start from the Navier–Stokes equations already discretised in
time and we will consider again a single time step, that is
uk+1 − uk k k+1
+ u ·∇ u − 2 div νD uk+1 + ∇pk+1 = f k+1 , in Ω, (14.13)
t
plus (14.1b) and (14.1c). Here, for the sake of simplicity (and without any loss of gen-
erality) we have chosen a semi-implicit treatment of the convective term. We wish now
to split the system in order to consider the effects of the velocity and the pressure terms
separately. We define an intermediate velocity ũ, obtained by solving the momentum
equation where the pressure contribution has been dropped, precisely
ũ − uk k
+ u ·∇ ũ − 2 div νD(ũ) = f k+1 , in Ω, (14.14a)
t
ũ = 0, on ∂Ω. (14.14b)
We may recognise that (14.14a) is now a problem on the velocity only, which could
be re-interpreted as the time discretisation of a parabolic differential equation of the
following type:
∂ ũ
+ (w·∇)ũ − 2 div νD(ũ) = f,
∂t
with w a given vector field. At this stage, we cannot impose the incompressibility con-
dition because we would obtain an over-constrained system.
We then consider the contribution given by the pressure term and the incompressibil-
ity constraint, that is
uk+1 − ũ
+ ∇pk+1 = 0, in Ω, (14.15a)
t
div uk+1 = 0, in Ω. (14.15b)
System (14.15) depends on both the velocity and pressure, yet we may derive an equa-
tion only for the pressure by taking (formally) the divergence of (14.15a) and exploiting
the incompressibility constraint (14.15b). That is,
k+1
u − ũ 1
0 = div + ∇pk+1 = − div ũ + div ∇pk+1
t t
1
=− div ũ + pk+1 ,
t
by which we obtain a Poisson equation for the pressure in the form
1
pk+1 = div ũ, in Ω. (14.16)
t
Eq. (14.16) must be supplemented by boundary conditions, which are not directly
available from the original problem (14.13). For that, we need to resort to the following
theorem, also known as Ladhyzhenskaja theorem.
T HEOREM 14.1 (Helmholtz decomposition principle). Let Ω be a domain of RN with

smooth boundary. Any vector function v ∈ L2 (Ω) (with N = 2, 3) can be uniquely rep-
resented as v = w + ∇ψ with w ∈ Hdiv (Ω), where

Hdiv (Ω) = w: w ∈ L2 (Ω), div w = 0, a.e. w · n = 0 on ∂Ω ,
and ψ ∈ H 1 (Ω).
S ECTION 14 51
The proof is rather technical and is here omitted. An outline, valid for the case v ∈
H1 (Ω), is given in C HORIN and M ARSDEN [1990]. A more general demonstration is
found in T EMAM [1984], Theorems 1 and 5.
If we now consider the expression

ũ = uk+1 + ∇ tpk+1 , (14.17)
derived from (14.15a), we may identify ũ with v and ( tpk+1 ) with ψ in the Helmholtz
decomposition principle. Then, the natural space for uk+1 is Hdiv(Ω), by which we
should impose
uk+1 · n = 0, on ∂Ω. (14.18)

Unfortunately, (14.18) is still a condition on the velocity, while we are looking for a
boundary condition for the pressure. The latter is found by considering the normal com-
ponent of (14.17) on the boundary,
ũ · n = uk+1 · n + t∇pk+1 · n, on ∂Ω,

and noting that on ∂Ω we have ũ · n = 0, because of (14.14b), and uk+1 · n = 0. Then,
∂pk+1
∇pk+1 · n = = 0, on ∂Ω,
∂n
which is a homogeneous Neumann boundary condition for the Poisson problem (14.16).
The projection method here presented for the solution of the Navier–Stokes equations
consists then in solving at each time-step a sequence of simpler problems, listed in the
following:
(1) Advection–diffusion problem for the velocity ũ. Solve problem (14.14a)–(14.14b).
(2) Poisson problem for the pressure
1
pk+1 = div ũ, in Ω, (14.19a)
t
∂ k+1
p = 0, on ∂Ω. (14.19b)
∂n
(3) Computation of uk+1 (this is an explicit step)
uk+1 = ũ − t∇pk+1 . (14.20)

For an analysis of projection methods as well as the set-up of higher order schemes
the reader may consult P ROHL [1997] and G UERMOND [1999]. We point out that pro-
jection schemes may also be used in conjunction with the semi-Lagrangian treatment of
the convective term (ACHDOU and G UERMOND [2000]).
14.5. Algebraic factorisation methods
An alternative way of reducing the computational cost of the solution of the full Navier–
Stokes problem is to operate at algebraic level. We will consider the generalised Stokes
problem in its algebraic form (14.7). This is the typical system that arises at each time
step of a time advancing scheme for the solution of the Navier–Stokes by a finite element
method, when the convective term is treated explicitly. In this case, the matrix C has the
form
M
C= + K + B,
t
where M is the mass matrix, K the stiffness matrix and B the matrix arising from the
explicit treatment of the convective term.
The matrix D derives from the discretisation of the divergence term, while D T rep-
resents a discrete gradient operator. We may formally solve for W,

W = C −1 Fs − D T Π , (14.21)
and by substituting into (14.7), we have
DC −1 D T Π = DC −1 Fs . (14.22)
The matrix DC −1 D T is called Stokes pressure matrix and is somehow akin to a discrete
Laplace operator. Having obtained Π from (14.22), we can then compute the velocity
by solving (14.21).
However, the inversion of C is in general prohibitive in terms of memory and com-
putational cost (indeed, C is sparse, but C −1 is not).
A way to simplify the computation can be found by recognising that steps (14.22) and
(14.21) may be derived from the following LU factorisation of the global matrix A:

C DT C 0 IW C −1 D T
A= = = LU, (14.23)
D 0 D −DC −1 D T D IΠ
where IW and IΠ indicate the identity matrices of dimension equal to the number of ve-
locity and pressure degrees of freedom, respectively. We then consider the LU solution

CW = Fs , W + C −1 D T Π = W,
−1
D W − DC D Π = 0,
T Π = Π,
where W and Π are intermediate velocities and pressures.
The scheme may be written in the following alternative form:
Intermediate velocity = Fs ,
CW (14.24a)
Pressure computation −DC −1
D Π = −D W,
T
(14.24b)
Velocity update − C −1 D T Π.
W=W (14.24c)
The key to reduce complexity is to replace C by a matrix simpler to invert, which, how-
ever, is “similar” to C, in a sense that we will make precise. This technique is called
inexact factorisation. In practise, we replace A in (14.23) by an approximation A∗ ob-
tained by replacing in the LU factorisation the matrix C −1 by convenient approxima-
tions, which we indicate by H1 and H2 , that is

C 0 IW H2 D T
A∗ = L∗ U ∗ =
D −DH1 D T D IΠ
T

C CH2 D
= . (14.25)
D D(H2 − H1 )D T
S ECTION 14 53
If we choose H2 = H1 , the discrete continuity equation is unaltered, that means that the
approximated system still guarantees mass conservation at discrete level. If H2 = C −1 ,
the discrete momentum equations are unaltered, and the resulting scheme satisfies the
discrete conservation of momentum. In particular, we can consider the two special cases

C CH D T
H1 = H2 = H ⇒ A∗ = ,
D 0

C DT
H1 = C −1 = H2 ⇒ A∗ = , Q = D H1 − C −1 D T .
D Q
14.5.1. The algebraic Chorin–Temam scheme

We note that
M 1 1
C= +K +B = M + t (K + B) = M IW + tM −1 (K + B) .
t t t
We recall the well-known Neumann expansion formula (M EYER [2000])
∞

(I + εA)−1 = (−1)j (εA)j ,
j =0
which converges for any matrix A and any positive number ε small enough to guarantee
that the spectral radius of εA is strictly less than one. We can apply this formula to C −1
to get
−1
C −1 = t IW + tM −1 (K + B) M −1
∞
j
= t (−1)j tM −1 (K + B) M −1
j =0

= t IW − tM −1 S + · · · M −1 , (14.26)
where we have put S = K + B.
A way to find a suitable approximation is to replace C −1 with a few terms of the se-
ries. The simplest choice considers just a first order approximation, which corresponds
to put into (14.25)
H1 = H2 = H = tM −1 . (14.27)
Consequently,

C tCM −1 D T C D T + tSM −1 D T
A∗ = ACT = = . (14.28)
D 0 D 0
The scheme obtained by applying the corresponding LU decomposition reads:
CW (14.29a)
Pressure computation − tDM D Π = −D W,
−1 T
(14.29b)
Velocity update W=W − tM −1 D T Π. (14.29c)
This algorithm is known as algebraic Chorin–Temam scheme. Comparing with the stan-
dard projection method, we may note that the algebraic scheme replaces in the pressure
computation step (14.29b) the Laplace operator of the Poisson problem (14.19) with a
“discrete Laplacian” DM −1 D T , which incorporates the boundary condition of the orig-
inal problem. No additional boundary condition is required for the pressure, contrary to
the standard (differential type) scheme.
R EMARK 14.1. The finite element mass matrix M is sparse and with the same graph
structure as C. Therefore, it may seem that there is little gain in the computational
efficiency with respect to the original factorisation (14.24). However, the matrix M may
be approximated by a diagonal matrix called lumped mass matrix (Q UARTERONI and
VALLI [1994]), whose inversion is now trivial.
R EMARK 14.2. It is possible to write the algebraic Chorin–Temam scheme in incre-

mental form, as it has been done for its differential counterpart.
14.5.2. The Yosida scheme

If we make the special choice
H1 = tM −1 , H2 = C −1 , (14.30)
we have

∗ C DT
A = AY = with Q = −D tM −1 − C −1 D T . (14.31)
D Q
The corresponding scheme reads
CW (14.32a)

Pressure computation − tDM D Π = −D W,−1 T
(14.32b)
Velocity update W=W − tC D Π.
−1 T
(14.32c)
The last step (14.32c) is more expensive than its counterpart (14.29c) in the Chorin–
Temam scheme, since now we need to invert the full matrix C. An analysis of this
method is found in Q UARTERONI , S ALERI and V ENEZIANI [1999].
R EMARK 14.3. If we consider the Stokes problem, we have C = ( t)−1 M + K and

consequently the matrix Q = −D( tM −1 − C −1 )D T in (14.31) may be written as

Q = − tD IW − (IW + tK)−1 D T = −( t)2 DY D T ,
where
1
Y= IW − (IW + tK)−1 ,
t
may be regarded as the Yosida regularisation of K, which is the discretisation of the
Laplace operator. That is Q may be interpreted as the discretisation of the differential
S ECTION 14 55
operator
( t)2 div(Y t ∇),
where Y t is the Yosida operator (B REZIS [1983]).
R EMARK 14.4. An incremental form may be found as follows. If Π n represents the

known value of the pressure degrees of freedom from the previous time step, we have
= Fs − D T Π n ,
Intermediate velocity C W

Pressure increment − tDM −1 D T Π − Π n = −D W,

Velocity update W=W − tC −1 D T Π − Π n .
More details on algebraic fractional step methods may be found in P EROT [1993] and
Q UARTERONI , S ALERI and V ENEZIANI [2000].
A major advantage of the algebraic factorisation schemes with respect to projection
methods is that they do not require to devise special boundary conditions for the pressure
problem, a task which is not always trivial.
All the techniques here presented may be extended to moving domains using the
procedure that will be illustrated in Section 18. In a moving domain context the various
matrices of the final algebraic system have to be recomputed at each times step to reflect
the change of domain geometry. As a consequence, a fully implicit approach is even
less computationally attractive, and factorisation schemes (at differential or algebraic
level), possibly with a semi-Lagrangian treatment of the convective term, are normally
preferred.
In the context of haemodynamics, algebraic factorisation schemes are particularly
attractive because of their flexibility with respect to the application of boundary condi-
tions. In particular, they can easily accommodate defective boundary conditions (F OR -
MAGGIA , G ERBEAU , N OBILE and Q UARTERONI [2002]).
C HAPTER IV
15. Mathematical modelling of the vessel wall
The vascular wall has a very complex nature and devising an accurate model for its me-
chanical behaviour is rather difficult. Its structure is indeed formed by many layers with
different mechanical characteristics (F UNG [1993], H OLZAPFEL , G ASSER and O GDEN
[2000]) (see Fig. 15.1). Moreover, experimental results obtained by specimens are only
partially significant. Indeed, the vascular wall is a living tissue with the presence of
muscular cells which contribute to its mechanical behaviour. It may then be expected
that the dead tissue used in the laboratory will have different mechanical characteristics
than the living one. Moreover, the arterial mechanics depend also on the type of the
surrounding tissues, an aspect almost impossible to reproduce in a laboratory. We are
then facing a problem whose complexity is enormous. It is the role of mathematical
modelling to find reasonable simplifying assumptions by which major physical charac-
teristics remain present, yet the problem becomes amenable to numerical analysis and
computational solution.
The set up of a general mathematical model of the mechanics of a solid continuum
may follow the same general route that we have indicated for fluid mechanics. In par-
ticular, it is possible to identify again a Cauchy stress tensor T. The major difference
between solids and fluids is in the constitutive relation which links T to kinematics
F IG . 15.1. The vessel wall is formed by many layers made of tissues with different mechanical character-
istics. Image taken from “Life: the Science of Biology” by W.K. Purves et al., fourth edition, published by
Sinauer Associates Inc. and W.H. Freeman and Company.
57
58 A. Quarteroni and L. Formaggia C HAPTER IV
quantities. We have seen in Section 5 that for a fluid such a kinematic quantity is the
velocity gradient or, more precisely, the strain rate D. For a solid, the Cauchy stress
tensor is instead a function of the deformation gradient, which we have already defined
in (7.7). That is, the constitutive law for a solid may be written as
T = T(Ft ).
If we assume that both the deformation gradient and the displacements are small, un-
der the hypothesis of linear elasticity and homogeneous material it is possible to de-
rive relatively simple relations for T. For sake of space, we will not pursue that matter
here. The interested reader may consult, for instance, the book by L.A. Segel (S EGEL
[1987], Chapter 4), or, for a more extensive treatment, the book by P.G. Ciarlet (C IAR -
LET [1988]).
Another possible situation is the one that involves a constitutive law of the form
T = T(D, Ft ), (15.1)
which describes the mechanical behaviour of a material with characteristics intermedi-
ate to those of a liquid and a solid. In such case, the continuum is said to be viscoelastic.
An example of such behaviour is given by certain plastics or by liquid suspensions. In
particular, also blood exhibits a viscoelastic nature, particularly when flowing in small
vessels, e.g. in arterioles and capillaries. Indeed, in that case the presence of suspended
particles and their interaction during the motion strongly affect the blood mechanical
behaviour. Again, we will not cover this topic here. The book by Y.C. Fung (F UNG
[1993]) may be used by the reader interested on the peculiar aspects of the mechanics
of living tissues.
The geometry of a section of an artery where no branching is present may be de-
scribed by using a curvilinear cylindrical coordinate system (r, θ, z) with the corre-
sponding base unit vectors er , eθ and ez , where ez is aligned with the axis of the artery,
as shown in Fig. 15.2.
Clearly, the vessel structure may be studied using full three-dimensional models,
which may also account for its multilayer nature. However, it is common practice to
F IG . 15.2. A model of a “realistic” section of an artery with the principal geometrical parameters.
S ECTION 16 59
F IG . 15.3. Different models for arterial wall mechanics.
resort to simplified 2D or even 1D mechanical models in order to reduce the overall

computational complexity when the final aim is to study the coupled fluid–structure
problem. In Fig. 15.3 we sketch some of the approximations normally made. A 2D
model may be obtained by either resorting to a shell-type description or considering
longitudinal sections (θ = const.) of the vessels. In the first case we exploit the fact that
the effective wall thickness is relatively small to reduce the whole structure to a surface.
A rigorous mathematical derivation (for the linear case) may be found in C IARLET
[1998]. In the second case we neglect the variations of the stresses in the circumferen-
tial direction. In this way we are able to eliminate all terms containing derivatives with
respect to θ in the equations and we may consider each plane θ = const. independently.
The resulting displacement field will depend only parametrically on θ . If, in addition,
we assume that the problem has an axial symmetry (which implies the further assump-
tion of a straight axis) the dependence on θ is completely neglected. In this case, also
the fluid would be described by a 2D axi-symmetric model.
The simplest models, called 1D models, are derived by making the same assumption
on the wall thickness made for the shell model, yet starting from a 2D model. The
structure will then be represented by a line on a generic longitudinal section, as shown
in the last picture of Fig. 15.3.
Even with all these simplifying assumptions an accurate model of the vessel wall
mechanics is rather complex. Therefore, in these notes we will only present the simplest
models, whose derivation is now detailed.
16. Derivation of 1D models of vessel wall mechanics
We are going to introduce a hierarchy of 1D models for the vessel structure of variable
complexity. We first present the assumptions common to all models.
The relatively small thickness of the vessel wall allows us to use as basis model a
shell model, where the vessel wall geometry is fully described by its median surface,
see Fig. 16.1.
F IG . 16.1. A cylindrical model of the vessel geometry. The latter is approximated, at any time t, by a surface
r = R(θ, z; t), which is outlined with dashed lines in figure.
We take as reference configuration Γ0w the one assumed by the vessel at rest when
filled with fluid with zero velocity and whose pressure is equal to the pressure Pext
exerted by the tissues external to the vessel. Although in principle Pext can change along
the vessel (for instance, because of the effect of gravity), for the sake of simplicity (and
without any loss of generality) we will consider only the case where Pext is constant.
The cylindrical-like aspect of sections of the arterial system allows us to derive sim-
plified mathematical models for the movement of the arterial wall assuming a straight
cylindrical geometry. We thus assume that the reference configuration Γ0w be a cylin-
drical surface with radius R0 (a regular strictly positive function of z), i.e.,

Γ0w = (r, θ, z): r = R0 (z), θ ∈ [0, 2π), z ∈ [0, L] ,
where L indicates the length of the arterial element under consideration. In our cylin-
drical coordinate system (r, θ, z), the z coordinate is aligned along the vessel axes and
a plane z = z̄ (= constant) defines an axial section.
We assume that the displacement vector η has only a radial component, that is
η = ηer = (R − R0 )er , (16.1)

where R = R(θ, z; t) is the function that provides, at each t, the radial coordinate r =
R(θ, z; t) of the wall surface. The current configuration Γtw at time t of the vessel
surface is then given by

Γtw = (r, θ, z): r = R(θ, z; t), θ ∈ [0, 2π), z ∈ [0, L] .
As a consequence, the length of the vessel does not change with time. We will indicate
with n the outwardly oriented unit normal to the surface Γtw at a given point. In Fig. 16.2
we sketch the reference and current configuration for the model of the section of an
artery.
Another important assumption is that of plain stresses. We neglect the stress com-
ponents along the normal direction n, i.e., we assume that the stresses lie on the vessel
surface.
We itemise here the main assumptions:
S ECTION 16 61
F IG . 16.2. The reference configuration Γ0w used for the derivation of our models is that of a circular cylinder.
Γtw indicates the current configuration at a given time t, while Ωt is the domain occupied by the fluid.
(A1) Small thickness and plain stresses. The vessel wall thickness h is sufficiently
small to allow a shell-type representation of the vessel geometry. In addition,
we will also suppose that it is constant in the reference configuration. The vessel
structure is subjected to plain stresses.
(A2) Cylindrical reference geometry and radial displacements. The reference vessel
configuration is described by a circular cylindrical surface with straight axes.3
The displacements are only in the radial direction.
(A3) Small deformation gradients. We assume that the deformation gradients are
small, so that the structure basically behaves like a linear elastic solid and
∂R/∂θ and ∂R/∂z remain uniformly bounded during motion.
(A4) Incompressibility. The vessel wall tissue is incompressible, i.e., it maintains its
volume during the motion. This is a reasonable assumption since biological
tissues are indeed nearly incompressible.
The models that we are going to illustrate could be derived from the general laws of
solid mechanics. Yet, this is not the route we will follow, preferring to describe them
in a more direct way, while trying to give some insight on the physical meaning of the
various terms that we are about to introduce.
16.1. Forces acting on the vessel wall
Let us consider the vessel configuration at a given time t and a generic point on the
vessel surface of coordinates θ = θ̄ , z = z̄ and r = R(θ̄ , z̄; t), with z̄ ∈ (0, L) and
θ ∈ (0, 2π). In the following derivation, if not otherwise indicated, all quantities are
computed at location (R(θ̄ , z̄; t), θ̄ , z̄) and at time t.
We will indicate with dσ the measure of the following elemental surface:

dθ dθ dz dz
dS = (r, θ, z): r = R(θ, z; t), θ ∈ θ̄ − , θ + , z ∈ z̄ − , z̄ + .
2 2 2 2
3 This assumption may be partially dispensed with, by assuming that the reference configuration is “close”
to that of a circular cylinder. The model here derived may be supposed valid also in that situation.
F IG . 16.3. A cylindrical model of the vessel geometry (left) and the infinitesimal portion of vessel wall used
for the derivation of the equations (right).
In Fig. 16.3 we have also indicated the two main stresses, the circumferential stress
and the longitudinal stress σ θ and σ z , which represent the internal forces acting on the
portion under consideration.
We may derive the following expression for n and dσ :

∂R ∂R
n = (R0 g)−1 Rer − eθ − R ez , (16.2)
∂θ ∂z
dσ = gR0 dθ dz = g dσ0 , (16.3)
where

2 2
R 1 ∂R ∂R
g= 1+ + ,
R0 R ∂θ ∂z
and σ0 = R0 dθ dz is the measure of the image of dS in the reference configuration Γ0w .
In particular, we have
R −1
n · er = g (16.4)
R0
and
n · er dσ = R dθ dz. (16.5)
The linear dimension of the elemental surface dS along the longitudinal direction has
been indicated with dl. It can be easily verified that

∂R 2
dl = 1 + dz. (16.6)
∂z
Let us now consider the external forces acting through the elemental surface dS.
S ECTION 16 63
• Forces from the surrounding tissues. As the tissue surrounding the vessel interacts
with the vessel wall structure by exerting a constant pressure Pext , the resulting
force acting on dS is simply given by
ftissue = −Pext n dσ + o(dσ ). (16.7)
• Forces from the fluid. The forces the fluid exerts on the vessel wall are represented
by the Cauchy stresses on the wall. Then, if we indicate with Tf the Cauchy stress
tensor for the fluid, we have
ffluid = −Tf · n dσ + o(dσ ) = P n dσ − 2µD(u) · n dσ + o(dσ ). (16.8)
16.2. The independent ring model
The independent ring model is expressed by a differential equation for the time evo-
lution of η, for each z and θ . For the derivation of this model, we will make some
additional assumptions:
(IR-1) Dominance of circumferential stresses σ θ . The stresses σ z acting along lon-
gitudinal direction are negligible with respect to σ θ and are thus neglected
when writing the momentum equation.
(IR-2) Cylindrical configuration. The vessel remains a circular cylinder during mo-
tion, i.e., ∂R/∂θ = 0. This hypothesis may be partially dispensed with, by
allowing small circumferential variations of the radius, yet we will neglect
∂R/∂θ in our model.
(IR-3) Linear elastic behaviour. Together with hypotheses (IR-1) and (IR-2) it al-
lows us to write that the circumferential stress is proportional to the relative
circumferential elongation, i.e.,
E η
σθ = , (16.9)
1 − ξ R0
2
where ξ is the Poisson ratio (which may be taken equal to 0.5 thanks to the
hypothesis (A4)) and E is the Young modulus.4
We will write the balance of momentum along the radial direction by analysing the
system of forces acting on dS. We have already examined the external forces, we need
now to look in more details at the effect of the internal forces, which, by assumption,
are only due to the circumferential stress σ θ .
We may note in Fig. 16.4 that the two vectors

dθ dθ
eθ θ̄ + and eθ θ̄ −
2 2
form with er an angle of π/2 + dθ/2 and −(π/2 + dθ/2), respectively. The component
of the resultant of the internal forces on the radial direction is then

dθ dθ
fint = σθ eθ θ̄ + + σθ eθ θ̄ − · er h dl
2 2
4 The presence of the term 1 − ξ 2 is due to the assumption of planar stresses. Some authors (like F UNG
[1984]) consider that the hypothesis of mono-axial stresses is more realistic for the problem at hand. In that
case one has to omit the term 1 − ξ 2 from the stress–strain relation and write simply σθ = EηR0−1 .
F IG . 16.4. Computation of the angle between σ θ and the radial direction er .
dθ
= −2σθ sin h dl = −σθ h dθ dl + o(dθ dl). (16.10)
2
Owing to the incompressibility assumption (A4), the volume in the current configura-
tion is unchanged with respect to that in the reference configuration, i.e.,
hR dθ dl = h0 R0 dθ dz.
Then, being o(dl) = o(dz), we may write (16.10) as
σθ Eh0 η
fint = − h0 R0 + o(dθ dz) = dθ dz + o(dθ dz).
R 1 − ξ2 R
Finally, the mass of the portion of vessel wall under consideration is
mass = ρw hR dθ dl = ρw h0 R0 dθ dz,
where ρw is the density of the vessel tissue, whereas the acceleration along the radial
direction is given by
∂ 2R ∂ 2η
= 2.
∂t 2 ∂t
By balancing the resultant of the internal and external forces, provided in (16.7) and
(16.8), with the inertia term, we have
∂ 2η Eh0 η
ρw h0 R0 2 dθ dz + dθ dz
∂t 1 − ξ2 R

= − 2µD(u) · n · er dσ + (P − Pext )n · er dσ + o(dθ dz). (16.11)
By dividing either side by dθ dz and passing to the limit for dθ → 0 and dz → 0, and
recalling that dσ = gR0 dθ dz = R(n·er )−1 dθ dz, thanks to (16.2) and (16.4), we obtain
∂ 2η Eh0 η
ρw h0 R0 + = − 2µD(u) · n · er gR0 + (P − Pext )R.
∂t 2 1−ξ R
2
Since the derivation has been made by considering an arbitrary plane θ = θ̄ and
time t, we may finally obtain the independent ring model
∂ 2η
+ bη = H, in Γ0w , t ∈ I, (16.12)
∂t 2
S ECTION 16 65
where
E
b= , (16.13)
ρw (1 − ξ 2 )R02
is a positive coefficient linked to the wall mechanical properties, while

1 R
H= (P − Pext ) − 2gµ D(u) · n · er
ρw h0 R0

ρ R
= (p − pext ) − 2gν D(u) · n · er , (16.14)
ρw h0 R0
is the forcing term which accounts for the action of external forces.
R EMARK 16.1. Often, the term R/R0 in the right-hand side of (16.14) is neglected as
well as the contribution to the forcing term due to the fluid viscous stresses. In this case,
we have just
P − Pext
H= (16.15)
ρw h0
and the forcing term does not depend anymore on the current geometrical configuration.
By neglecting the acceleration term in (16.12), we obtain the following algebraic

model, which is often found in the medical and bioengineering literature:
bη = H, in Γ0w , t ∈ I, (16.16)
according to which the wall displacement is proportional to the normal component of
the applied external stresses.
R EMARK 16.2. One may account for the viscoelastic nature of the vessel wall structure
even in this simple model by adding to the constitutive relation (16.9) a term propor-
tional to the displacement velocity, as in a simple Voigt–Kelvin model (F UNG [1993]),
that is by writing
E η γ ∂η
σθ = + ,
1 − ξ 2 R0 R0 ∂t
where γ (whose unit is [γ ] = kg/m s) is a positive constant damping parameter.
Then, the resulting differential equation would read:
∂ 2η γ ∂η
+ 2 + bη = H, in Γ0w , t ∈ I. (16.17)
∂t 2 R0 ρw h0 ∂t
We may note that the term R10 ∂η ∂t plays the role of the strain rate D into the general
relation for viscoelastic materials (15.1).
Models (16.12), (16.16) and (16.17) are all apt to provide a solution η for every
possible value of θ . In principle, since no differentiation with respect to θ is present in
the model, nothing would prevent us to get significant variations of η with θ (or even
a discontinuity), which would contradict assumption (IR-2). This potential drawback
could be eliminated by enriching the models with further terms involving derivatives
along θ , as in the case of models derived from shell theory (C IARLET [2000]). On the
other hand, a more heuristic and less rigorous argument can be put forward moving
from (16.16). Since b is relatively large, smooth variations of the forcing term H with
respect to θ are damped to tiny one on η. This observation may be extended also to
models (16.12) and (16.17) in view of the fact that for the problems at hand the term
bη dominates the other terms on the left-hand side. Similar considerations apply to the
model that we will introduce in the next subsection.
16.3. The generalised string model
A more complete model (Q UARTERONI , T UVERI and V ENEZIANI [2000]) considers

also the effects of the longitudinal stresses σ z . Experimental and physiological analysis
(F UNG [1993]) show that vessel walls are in a “pre-stressed” state. In particular, when
an artery is extracted from a body tends to “shrink”, i.e., to reduce its length. This fact
implies that arteries in the human body are normally subjected to a longitudinal tension.
At the base of the generalised string model is the assumption that this longitudinal
tension is indeed the dominant component of the longitudinal stresses.
More precisely, let us refer to Fig. 16.5; we replace assumption (IR-1) by the follow-
ing:
(GS-1) The longitudinal stress σ z is not negligible and, in particular,
σ z = ±σz τ , (16.18)
where τ is the unitary vector tangent to the curve
r = R(θ̄ , z; t), (16.19)
and its modulus σz is constant. Moreover, we assume that it is a traction stress
(that is with a versus equal to that of the normal to the surface on which it
applies).
We also maintain assumption (IR-2) of the independent ring model. When consider-
ing the forces acting on dS, we have now a further term, namely (referring again to
Fig. 16.5)

fz = σ z (z̄ + dz/2) + σ z (z̄ − dz/2) hR dθ
τ (z̄ + dz/2) − τ (z̄ − dz/2) dτ
= σz dlhR dθ = σz R0 h0 dl dθ + o(dz dθ ).
dl dl
We now exploit the Frenet–Serret formulae to write
dτ
= κn,
dl
where κ is the curvature of the line r = R(θ̄ , z; t), whose expression is

∂ 2R ∂R 2 −3/2
κ = 2 1+ . (16.20)
∂z ∂z
S ECTION 16 67
F IG . 16.5. A cylindrical model of the vessel geometry (top) and quantities on a longitudinal section (bottom).
By recalling (16.6) and (16.4), we obtain

∂ 2R ∂R 2 −3/2
fz · er = σz 2 1 + R0 h0 dz dθ + o(dz dθ ).
∂z ∂z
We eliminate the geometric non-linearity in the model by neglecting the term (∂R/∂z)2 .
Furthermore, we replace ∂ 2 R/∂z2 by ∂ 2 η/∂z2 .5
By proceeding like in the previous section, we may modify the independent ring
model into the following differential equation:
∂ 2η ∂ 2η
2
− a 2 + bη = H, in Γ0w , t ∈ I, (16.21)
∂t ∂z
where
σz
a= .
ρw h0
5 This last equality is clearly true whenever R is varying linearly with z.
0
The final generalised string model is obtained by adding to the expression for σ z in
(16.18) a term
∂ ∂η
c , c > 0,
∂t ∂z
which is a viscoelastic term linking the longitudinal stress to the rate of rotation of the
structure. For small displacements, ∂η/∂z is indeed proportional to the angle of rota-
tion around the circumferential direction of the structure, with respect to the reference
configuration.
The result is
∂ 2η ∂ 2η ∂ 3η
− a + bη − c = H, in Γ0w , t ∈ I. (16.22)
∂t 2 ∂z2 ∂t∂z2
17. Analysis of vessel wall models
In the following we will provide some a-priori estimates for the differential models just
proposed.
We recall Poincaré inequality for the one-dimensional case.
L EMMA 17.1 (Poincaré inequality – one-dimensional case). Let f ∈ H 1 (a, b) with

f (a) = 0. Then there exists a positive constant Cp such that

df
f L2 (0,L) Cp
dx 2 . (17.1)
L (0,L)
P ROOF. For all x ∈ [a, b] we have,

x x
df df
f (x) = f (a) + (τ ) dτ = (τ ) dτ.
a dx a dx
Then,
b b s 2
df
f 2 (s) ds = (τ ) dτ ds
a a a dx
b s 1/2 s 2 1/2 2
df
2
1 dτ (τ ) dτ ds
a a a dx
(by Cauchy–Schwarz inequality)
b 2 2
df
2 df
(b − a)
dx 2 ds = (b − a) dx 2 ,
a L (a,b) L (a,b)
by which inequality (17.1) is proved by taking Cp = (b − a). The same inequality holds
if f (b) = 0.
Thanks to the fact that no derivatives with respect to the variable θ are present in the
equations, we may carry out some further analysis of the structure models illustrated so
far by considering the equations for a fixed value of θ and z.
S ECTION 17 69
We will consider Eq. (16.12) and address then the following problem:
∂ 2η
+ bη = H, in Γ0w , t ∈ I, (17.2)
∂t 2
with the following initial values for the displacement and its time rate:
∂η
η = η0 , = η1 , in Γ0w , t = t0 . (17.3)
∂t
We also introduce the space L2 (I ; L2 (Γ0w )) of functions f : Γ0w × I → R that are
square integrable in Γ0w for almost every (a.e.) t ∈ I and such that
t1

f (τ )2 2 w dτ < ∞.
L (Γ ) 0
t0
L EMMA 17.2. If H ∈ L2 (I ; L2 (Γ0w )), the following inequality holds for a.e. t ∈ I :

∂η 2 2
(t)
∂t 2 w + b η(t) L2 (Γ0w )
L (Γ0 )
t

η1 2L2 (Γ w ) + bη0 2L2 (Γ w ) + H (τ )2 2 w dτ e(t −t0) . (17.4)
0 0 L (Γ )
0
t0
P ROOF. It can be obtained by multiplying (17.2) by ∂η/∂t and applying Gronwall

lemma (Lemma 10.3).
Relation (17.4) asserts that the sum of the total kinetic and elastic potential energy
associated to Eq. (17.2) is bounded, at each time t, by a quantity which depends only
on the initial condition and the forcing term.
Let us consider the generalised string model (16.22) with the following initial and
boundary conditions:
∂η
η = η0 , = η1 in Γ0w , t = t0 , (17.5a)
∂t
η|z=0 = α, η|z=L = β, t ∈ I. (17.5b)
Let us define the following energy function:
2
1
∂η (t)
∂η 2
(t)

η(t)2 2 w .
es (t) = + a + b (17.6)
2 ∂t L2 (Γ w ) ∂z 2 w
L (Γ )
L (Γ0 )
0 0
L EMMA 17.3. If H ∈ L2 (I ; L2 (Γ0w )) and α = β = 0, the following inequality holds

for a.e. t ∈ I :
2 2 t1
c t
∂ η (τ )

H (τ )2 2 w dτ,
es (t) + dτ e s (0) + k L (Γ0 )
(17.7)
2 t0 ∂t∂z L2 (Γ w ) t0
0
where k = Cp2 /(2c) and Cp is the Poincaré constant.

P ROOF. We use the short-hand notations η̇ and η̈ for the time derivatives of η. We first
multiply the generalised string equation (16.22) by η̇ and integrate w.r. to z:
L L 2 L L
∂ η ∂ 3η
η̇η̈ − a η̇ 2 − c η̇ +b η̇η
0 0 ∂z 0 ∂t∂z2 0
L 2 L 2 2
1 d L 2 ∂ η ∂η ∂η L ∂ η
= η̇ + a − a η̇ +c
2 dt 0 0 ∂t∂z ∂z ∂z 0 0 ∂t∂z
2 L L L
∂ η b d
−c η̇ + η2 = η̇H. (17.8)
∂t∂z 0 2 dt 0 0
By exploiting the homogeneous boundary conditions and the fact that

∂ 2 η ∂η 1 ∂ ∂η 2
= ,
∂t∂z ∂z 2 ∂t ∂z
we have
1 d
L a d
L ∂η 2
L ∂ 2η
2
b d
L L
η̇2 + +c + η2 = η̇H.
2 dt 0 2 dt 0 ∂z 0 ∂t∂z 2 dt 0 0
Thanks to the hypothesis of axial symmetry, we have
2 2
des ∂ η
+ c = η̇H. (17.9)
dt ∂t∂z L2 (Γ w ) Γ0w 0
The application the Cauchy–Schwarz, Young and Poincaré inequalities to the right-hand
side gives
2 2
des ∂ η 1
+ c

H 2L2 (Γ w ) + εη̇2L2 (Γ w )
dt ∂t∂z L2 (Γ w ) 4ε 0 0
0
2 2
1 ∂ η
H 2L2 (Γ w ) + Cp2 ε
∂t∂z 2 w
4ε 0
L (Γ ) 0
for any positive ε. If we choose ε such that Cp2 ε = c/2 and integrate in time between t0
and t, we finally obtain the desired result.
C HAPTER V
18. The coupled fluid structure problem
In this part we will treat the situation arising when the flow in a vessel interacts mechan-
ically with the wall structure. This aspect is particularly relevant for blood flow in large
arteries, where the vessel wall radius may vary up to 10% because of the forces exerted
by the flowing blood stream.
We will first illustrate a framework for the Navier–Stokes equations in a moving
domain which is particularly convenient for the analysis and for the set up of numerical
solution methods.
18.1. The Arbitrary Lagrangian Eulerian (ALE) formulation of the Navier–Stokes

equation
In Section 9 we have introduced the Navier–Stokes equations in a fixed domain Ω,

according to the Eulerian approach where the independent spatial variables are the co-
ordinates of a fixed Eulerian system. We now consider the case where the domain is
moving. In practical situations, such as the flow inside a portion of a compliant artery,
we have to compute the flow solution in a computational domain Ωt varying with time.
The boundary of Ωt may in general be subdivided into two parts. The first part coin-
cides with the physical fluid boundary, i.e., the vessel wall. In the example of Fig. 18.1,
this part is represented by Γtw , which is moving under the effect of the flow field. The
other part of ∂Ωt corresponds to “fictitious boundaries” (also called artificial bound-
aries) which delimit the region of interest. They are necessary because solving the fluid
equation on the whole portion of space occupied by the fluid under study is in general
F IG . 18.1. The longitudinal section of a model of an artery. The vessel wall Γtw is moving. The location
along the z axis of Γtin and Γtout are fixed.
71
72 A. Quarteroni and L. Formaggia C HAPTER V
impractical, if not impossible. In our case, that would mean solving the whole circula-
tory system!
In the example of Fig. 18.1, the artificial boundaries are the inlet and outlet bound-
aries, there indicated by Γtin and Γtout , respectively. The location of these boundaries is
fixed a priori. More precisely, Γtin and Γtout may change with time because of the dis-
placement of Γtw , however they remain planar and their position along the vessel axis
is fixed.
Clearly in this case the Eulerian approach becomes impractical.
A possible alternative would be to use the Lagrangian approach. Here, we identify
the computational domain on a reference configuration Ω0 and the corresponding do-
main in the current configuration, which we indicate with ΩLt , will be provided by the
Lagrangian mapping (which has been introduced in Section 7), i.e.,
ΩLt = Lt (Ω0 ), t ∈ I. (18.1)

Fig. 18.2 illustrates the situation for the flow inside an artery whose wall is moving.
Since the fluid velocity at the wall is equal to the wall velocity, the Lagrangian mapping
effectively maps Γ0w to the correct wall position Γtw at each time t. However, the “fic-
titious” boundaries Γ0in and Γ0out in the reference configuration will now be transported
along the fluid trajectories, into ΓLint and ΓLout
t
. This is clearly not acceptable, particu-
larly if one wants to study the problem for a relatively large time interval. Indeed, the
domain rapidly becomes highly distorted.
The ideal situation would then be that indicated in Fig. 18.2(b). Even if the wall
is moving, one would like to keep the inlet and outlet boundaries at the same spatial
location along the vessel axis.
F IG . 18.2. Comparison between the Lagrangian and the ALE approach. The reference computational domain
Ω0 is mapped by (a) the Lagrangian mapping Lt and by (b) the Arbitrary Lagrangian Eulerian mapping.
S ECTION 18 73
With that purpose, we introduce the Arbitrary Lagrangian Eulerian (ALE) mapping
At : Ω0 → ΩAt , Y → y(t, Y) = At (Y), (18.2)
which provides the spatial coordinates (t, y) in terms of the so-called ALE coordinates
(t, Y), with the basic requirement that At retrieves, at each time t ∈ I , the desired com-
putational domain, i.e.,
ΩAt ≡ At (Ω0 ) = Ωt , ∀t ∈ I.
The ALE mapping should be continuous and bijective in Ω0 . Once given, we may
define the domain velocity field as
∂

w(t, Y) = y(t, Y), (18.3)
∂t
which in the spatial coordinates is expressed as

w= w ◦ A−1
t , w t, A−1
i.e., w(t, y) = t (y) . (18.4)
Similarly to what has been done for the Lagrangian mapping in Section 7 we use the
convention of indicating by f˜ the composition of a function f with the ALE mapping,
i.e., f˜ = f ◦ At .
We define the ALE trajectory TY for every Y ∈ Ω0 as

TY = t, y(t, Y) : t ∈ I (18.5)
and the ALE derivative of a function f , which we denote by (D A /Dtf ), as the time
derivative along a trajectory TY , that is if
f : I × Ωt → R,
then
DA DA ∂ f˜
f : I × Ωt → R, f (t, y) = (t, Y), Y = A−1
t (y). (18.6)
Dt Dt ∂t
Similarly to what already obtained for the Lagrangian mapping (relation (7.4)), we
have
DA ∂f
f= + w ·∇ f, (18.7)
Dt ∂t
where now the gradient is made with respect to the y-coordinates.
The Jacobian of the ALE mapping J At , defined as

∂y
J At = det , (18.8)
∂Y
is, for all t ∈ I , a positive quantity because the ALE mapping is surjective and at time
t0 is equal to the identity mapping. It satisfies the following relation:
D A At
J = J At div w. (18.9)
Dt
Again in a way all analogous to what seen for the Lagrangian mapping, we may derive
the following result.
T HEOREM 18.1 (ALE transport theorem). Let V0 ⊂ Ω0 , and let V At ⊂ Ωt be its image
under the mapping At . Furthermore, let f : I × Ωt → R be continuously differentiable
with respect to both variables. Then
A
d D ∂f
f= f + f div w = + div(f w)
dt V At At Dt V At ∂t
V
∂f
= + f w · n. (18.10)
V At ∂t ∂V At
The proof is similar to that of Theorem 7.2 and is omitted.

The Navier–Stokes equations (9.1) are clearly valid on Ωt , yet it may be convenient
to recast them in order to put into evidence the ALE time derivative. We obtain, by a
straightforward application of (18.7) to (9.1),
DA
u + (u − w)·∇ u + ∇p − 2 div νD(u) = f,
Dt
div u = 0, (18.11)
in Ωt and for all t ∈ I .
18.2. Coupling with the structure model
We now study the properties of the coupled fluid–structure problem, using for the struc-
ture the generalised string model (16.22). Referring to Fig. 16.2, we recall that Γtw is
the current configuration of the vessel structure, while Γ0w is the reference configuration
in which the structure equation is written. We also recall that we take n always to be the
outwardly vector normal to the fluid domain boundary.
We will then address the following problem:
For all t ∈ I , find u, p, η such that
DA
u + (u − w)·∇ u + ∇p − 2 div νD(u) = f,
Dt
div u = 0, in Ωt , (18.12)
and
∂ 2η ∂ 2η ∂ 3η
− a + bη − c = H, in Γ0w (18.13)
∂t 2 ∂z2 ∂t∂ 2 z
with the following initial conditions for t = t0 :
u = u0 , x ∈ Ω0 , (18.14a)
η = η0 , η̇ = η1 , in Γ0w , (18.14b)
boundary conditions for t ∈ I ,

2νD(u) − (p − Pext )I · n = 0, on Γtout, (18.15a)
S ECTION 18 75
u = g, on Γtin , (18.15b)
η|z=0 = α, η|z=L = β, (18.15c)
and the interface condition
∂η
ũ = u ◦ At = er , on Γ0w , t ∈ I. (18.16)
∂t
Another interface condition is implicitly provided by the fact that the forcing term H
is function of the fluid variables (see (16.14)).
Here, u0 , g, α and β are given functions, H is the forcing term (16.14) and At is an
ALE mapping such that A−1 t (∂Ωt ) = Γ ∪ Γ
in out ∪ Γ w . We have used the ALE form
0
for the Navier–Stokes equations since it is best suited in view of the numerical solution,
as it will be detailed in the next section.
We may then recognise the sources of the coupling between the fluid and the structure
models, which are twofold (in view of a possible iterative solution strategy):
• fluid → structure. The fluid solution provides the value of H , which is function of
the fluid stresses at the wall.
• structure → fluid. The movement of the vessel wall changes the geometry on which
the fluid equations must be solved. In addition, the proper boundary conditions for
the fluid velocity in correspondence to vessel wall are not anymore homogeneous
Dirichlet conditions, but they impose the equality between the fluid and the struc-
ture velocity. They express the fact that the fluid particle in correspondence of the
vessel wall should move at the same velocity as the wall.
Note that we have made some changes with respect to the nomenclature used in (9.1)
to indicate that the domain is now moving. We rewrite the expression of the forcing
term H , given in (16.14), by noting that while the fluid velocity and pressure are writ-
ten in the current configuration, H lives in the reference configuration for the vessel
wall Γ0w . Therefore, following the nomenclature introduced in the previous subsection,
we write

ρ R
H= (p̃ − p0 )
− 2g ν̃ D(u) · n · er . (18.17)
ρw h0 R0
18.3. An energy inequality for the coupled problem
In this section we will obtain an a-priori inequality for the coupled fluid–structure prob-
lem just presented. We will consider only the case of homogeneous boundary condi-
tions, that is
g = 0, α = β = 0,
for the coupled problem (18.12)–(18.16).
L EMMA 18.1. The coupled problem (18.12)–(18.16) with g = 0 and α = β = 0 satis-

fies the following energy equality for all t ∈ I :
2 2
d ω
u(t) 2
∂ η
L (Ωt )
+ es (t) + 2ω νD(u) : D(u) + c
∂z∂t 2 w
dt 2 Ωt L (Γ ) 0

ω
+ |u|2 u · n = ω f · u, (18.18)
2 Γtout Ωt
where es was defined in (17.6) and

ρ
ω= . (18.19)
ρw h0
Moreover, if we assume that the net kinetic energy flux is non-negative on the outlet
section, i.e.,

|u|2 u · n 0 ∀t ∈ I, (18.20)
Γtout
we obtain the a-priori energy estimate

ω
u(t) 2
t
∇u(τ )2 2
L (Ωt )
+ es (t) + CK ων0 L (Ωτ )
dτ
2 t0
t 2 2
∂ η
+c ∂z∂t (τ )
2 0 dτ
t0 L (Γw )

ω ωCP2 t
u0 L2 (Ωt ) + es (t0 ) + f(τ )2L2 (Ω ) dτ, t ∈ I. (18.21)
2 4νCK t0
τ
P ROOF. We recall expression (17.9) and we recast the right-hand side on the current
configuration Γtw . By exploiting (16.3) and (16.4), we have

∂η ρ R ∂η
H dσ0 =
(p̃ − pext ) − 2gν D(u) · n · er dσ0
Γ0w ∂t ρ h
w 0 Γ0 w R 0 ∂t

∂η
=ω
(p̃ − pext )n · er − 2ν D(u) · n · er g dσ0
Γ0w ∂t

=ω
(p̃ − pext )n − 2ν D(u) · n · ũg dσ0
Γ0w

=ω (p − pext )n − 2ν D(u) · n · u dσ,
Γtw
where we have used the interface conditions (18.16). Then,

2 2
1 des ∂ η

+ c =ω (p − pext )n − 2ν D(u) · n · u dσ. (18.22)
2 dt
∂z∂t L2 (Γ w ) w
Γt 0
As for the fluid equations, we follow the same route of Theorem 12.1. In particular,
we begin by multiplying (18.12) by u and integrating over Ωt , obtaining

DA
u· u+ u · (u − w) · ∇ u + u · ∇p − 2ν div D(u) = (f, u).
Ωt Dt Ωt Ωt
(18.23)
S ECTION 18 77
We now analyse each term in turn. By exploiting the ALE transport theorem (18.10),
we may derive that

DA ∂
u 1 ∂|ũ|2
u· u= Jt ũ · = Jt
Ωt Dt Ω0 ∂t 2 Ω0 ∂t

1 DA 2 1 d 1
= |u| = |u|2 − |u|2 div w. (18.24)
2 Ωt Dt 2 dt Ωt 2 Ωt
The convective term gives

u · (u − w)·∇ u
Ωt

1 1 1
=− |u|2 div u + |u|2 div w + |u|2 (u − w) · n
2 Ωt 2 Ωt 2 ∂Ωt

1 1
= |u|2 div w + |u|2 u · n, (18.25)
2 Ωt 2 Γtout
since div u = 0 in Ωt while w = u on Γtw and w = 0 on ∂Ωt \ Γtw .
The other terms provide

u ·∇ p = (since pext = const.) u ·∇ (p − pext )
Ωt Ωt

= − (p − pext ) div u + (p − pext )u · n
Ωt ∂Ωt

= (p − pext )u · n + (p − pext )u · n (18.26)
Γtout Γtw
and

νu · div D(u) = − ν∇u : D(u) + νu · D(u) · n
Ωt Ωt ∂Ωt

=− νD(u) : D(u) + νu · D(u) · n
Ωt ∂Ωt

=− νD(u) : D(u) + ν D(u) · n · u
Ω Γtout
t

+ ν D(u) · n · u, (18.27)
Γtw
where we have exploited again the symmetry of D(u).

Using the results obtained in (18.24)–(18.27) into (18.23), rearranging the terms and
recalling the boundary condition (18.15a), we can write

1 d 1
u2L2 (Ω ) + 2 νD(u) : D(u) + |u|2 u · n
2 dt t
Ωt 2 Γtout

+ (p − pext )n − 2νD(u) · n · u = f · u.
Γtw Ωt
We now recall expression (18.22) and recognise the equivalence of the integrals
over Γtw , which express the exchange of power (rate of energy) between fluid and struc-
ture. We multiply then the last equality by ω and add it to (18.22), obtaining (18.18).
Using (18.20), (10.2) and the fact that ν ν0 > 0,
2 2
d ω
u(t) 2
∂ η

L (Ω )
+ e s (t) + 2C K ων0 ∇u 2
L 2 (Ω ) + c ∂z∂t 2 w
dt 2 t t
L (Γ0 )
2 2
d ω
u(t) 2
∂ η
L (Ωt )
+ es (t) + 2ω νD(u) : D(u) + c
∂z∂t 2 w
dt 2 Ωt L (Γ0 )

ω ω
ω f · u f2L2 (Ω ) + ωεu2L2 (Ω ) f2L2 (Ω ) + CP2 ωε∇u2L2 (Ω ) ,
Ωt 4ε t t 4ε t t
for any positive ε. To derive the last inequality we have applied the Poincaré inequality
(10.1).
The desired result is then obtained by taking ε = (ν0 CK )/CP2 and integrating in time
between t0 and t.
This last result shows that the energy associated to the coupled problem is bounded,
at any time, by quantities which depend only on the initial condition and the applied
volume forces. Moreover, since in blood flow simulation we neglect the volume force
term f in the Navier–Stokes equations, estimate (18.21) simplifies into
t
ω
u(t) 2

∇u(τ )2 2
L (Ωt )
+ e s (t) + 2C K ων0 L (Ωτ )
dτ
2 t0
t 2 2
∂ η
+c ∂z∂t (τ )
2 0 dτ
t0 L (Γw )
ω
u0 L2 (Ωt ) + es (t0 ), ∀t ∈ I.
2
R EMARK 18.1. We may note that the non-linear convective term in the Navier–Stokes
equations is crucial to obtain the stability result, because it generates a boundary term
which compensates that coming from the treatment of the velocity time derivative.
These two contributions are indeed only present in the case of a moving boundary.
R EMARK 18.2. Should we replace the boundary condition (18.15a) by

1
2νD(u) · n − p − pext + |u|2 n = 0 on Γtout, t ∈ I, (18.28)
2
we would obtain the stability results without the restrictions on the outlet velocity
(18.20).
Let us note that the above boundary condition amounts to imposing a zero value for
the total stress at the outflow surface.
R EMARK 18.3. Under slightly different assumptions, that is periodic boundary condi-
tions in space and the presence of a further dissipative term proportional to ∂ 4 η/∂z4 in
S ECTION 18 79
the generalised string model, B EIRÃO DA V EIGA [2004] has recently proven an exis-
tence result of strong solutions to the coupled fluid–structure problem. The well posed-
ness of fluid–structure interaction solutions in more general settings is still a largely
open problem. A review of recent theoretical results may be found in G RANDMONT
and M ADAY [2000].
The hypothesis (18.20) is obviously satisfied if Γtout is indeed an outflow section, i.e.,
u · n 0 for all x ∈ Γtout . As already pointed out, this is seldom true for vascular flow,
particularly in large arteries.
We may observe that the “viscoelastic term” −c(∂ 3η/(∂t∂ 2 z)) in (16.22) allows to
obtain the appropriate regularity of the velocity field u on the boundary (see N OBILE
[2001]).
In the derivation of the energy inequality (18.21), we have considered homogeneous
boundary conditions both for the fluid and the structure. However, the conditions η = 0
at z = 0 and z = L, which correspond to hold the wall fixed at the two ends, are not
realistic in the context of blood flow. Since the model (16.22) for the structure is of
propagative type, the first order absorbing boundary conditions
∂η √ ∂η
− a = 0 at z = 0, (18.29)
∂t ∂z
∂η √ ∂η
+ a = 0 at z = L (18.30)
∂t ∂z
look more suited to the problem at hand. An inequality of the type (18.21) could still be
proven. Indeed, the boundary term which appears in (17.8) would now read
2 2
∂η ∂η ∂ 2 η ∂η z=L √ ∂η ∂η
− a +c = a +
∂z ∂t ∂z∂t ∂t z=0 ∂t z=0 ∂t z=L
2 2
c√ d ∂η ∂η
+ a + .
2 dt ∂t z=0 ∂t z=L
This term, integrated in time, would eventually appear on the left-hand side of inequality
(18.21). We may note, however, that we obtain both for z = 0 and z = L the following
expression:
t 2 2 2
√ ∂η c 1 ∂η c 1 ∂η
a (τ ) dτ + (t) = (t0 ) . (18.31)
t0 ∂t 2 a ∂t 2 a ∂t
This additional term is positive and depends only on initial conditions.
Yet, conditions (18.29) and (18.30) are not compatible with the homogeneous Dirich-
let boundary conditions for the fluid; indeed, if η|z=0 = 0 and u = 0 on Γtin , the trace of
u on the boundary is discontinuous and thus not compatible with the regularity required
on the solution of (18.12) (see, e.g., Q UARTERONI and VALLI [1994]).
A possible remedy consists of changing the condition u = 0 on Γtin into
u · ez = gz ◦ A−1
t , (T · n) × ez = 0
on Γtin , where gz is a given function defined on Γ0in , with gz = 0 on ∂Γ0in . Here T is the
stress tensor defined in (8.13). An energy inequality for the coupled problem can be de-
rived also in this case with standard calculations, taking a suitable harmonic extension
g˜z of the non-homogeneous data gz . The calculations are here omitted for the sake of
brevity.
19. An iterative algorithm to solve the coupled fluid–structure problem
In this section we outline an algorithm that at each time-level allows the decoupling of
the sub-problem related to the fluid from that related to the vessel wall. As usual, t k ,
k = 0, 1, . . . denotes the kth discrete time level; t > 0 is the time-step, while v k is the
approximation of the function (scalar or vector) v at time t k .
The numerical solution of the fluid–structure interaction problem (18.12), (18.13)
will be carried out by constructing a suitable finite element approximation of each sub-
problem. In particular, for the fluid we need to devise a finite element formulation suit-
able for moving domains (or, more precisely, moving grids). In this respect, the ALE
formulation will provide an appropriate framework.
To better illustrate the situation, we refer to Fig. 19.1 where we have drawn a 2D
fluid–structure interaction problem. The fluid domain is Ωt and the movement of its
upper boundary Γtw is governed by a generalised string model. This geometry could be
derived from an axisymmetric model of the flow inside a cylindrical vessel. However, in
this case we should employ the Navier–Stokes equations in axisymmetric coordinates.
Since this example is only for the purpose of illustrating a possible set-up for a coupled
fluid–structure algorithm, for the sake of simplicity we consider here a two-dimensional
fluid–structure problem governed by Eqs. (18.12), (18.13), with interface conditions
(18.16), initial and boundary conditions (18.14) and the additional condition
u|Γ 0 = 0, t ∈ I.
The algorithm here presented may be readily extended to three-dimensional problems.
The structure on Γ0w will be discretised by means of a finite element triangula-
tion Ths , like the one we illustrate in Fig. 19.2. We have considered the space Sh of
piece-wise linear continuous (P1) finite elements functions to represent the approxi-
mate vessel wall displacement ηh . In the same figure we show the position at time t
of the discretised vessel wall boundary Γt,hw , corresponding to a given value of the dis-
crete displacement field ηh ∈ Sh . Consequently, the fluid domain will be represented

f
at every time by a polygon, which we indicate by Ωt,h . Its triangulation Tt,h will be
F IG . 19.1. A simple fluid–structure interaction problem.

S ECTION 19 81
F IG . 19.2. Position of the discretised vessel wall corresponding to a possible value of ηh .
F IG . 19.3. The triangulation used for the fluid problem at each time t is the image through a map At of a
mesh constructed on Ω0 .
f
constructed as the image by an appropriate ALE mapping At of a triangulation T0,h
of Ω0 , as shown in Fig. 19.3. Correspondingly, Ωt,h = At Ω0,h , where Ω0,h is the ap-
f
proximation of Ω0 induced by the triangulation Tt,h (clearly, if Ω0 has a polygonal
f
boundary we have Ω0,h = Ω0 .) The trace of T0,h on Γ0w will coincide with the “tri-
angulation” Ths of the vessel wall, thus we consider geometrically conforming finite
elements between the fluid and the structure. The possibility of using a geometrically
non-conforming finite element representation has been investigated in G RANDMONT
and M ADAY [1998].
We then have to face the following problem. Suppose that we know at t = t k+1 a
discrete displacement field ηhk+1 and thus the corresponding position of the domain
f
boundary ∂Ωt k+1,h . How to build a map At k+1 such that At k+1 (T0,h ) is an acceptable
finite element mesh for the fluid domain? This task is in general not simple. However,
if we can assume that Ωt,h is convex for all t and that the displacements are relatively
small, the technique known as harmonic extension may well serve the purpose. Let Xh
f
be the P1 finite element vector space associated to T0,h , while
X0h = {wh ∈ Xh : wh |∂Ω0,h = 0}

and let gh : ∂Ω0,h → ∂Ωt k+1,h be the function describing the fluid domain boundary.
We build the map by seeking yh ∈ Xh such that

∇yh : ∇zh = 0 ∀zh ∈ X0h , yh = gh , on ∂Ω0,h , (19.1)
Ω0
and then setting At k+1 (Y) = yh (Y), ∀Y ∈ Ω0,h . This technique has indeed been
adopted for the mesh in Fig. 19.3. From a practical point of view, the value of yh in
f f
correspondence to the nodes of T0,h gives the position of the corresponding node in Tt,h
at time t k+1 . A more general discussion on the construction of the ALE mapping may
be found in F ORMAGGIA and N OBILE [1999], N OBILE [2001] as well as in G ASTALDI
[2001].
R EMARK 19.1. Adopting P1 elements for the construction of the ALE map ensures
f f
that the triangles of Th,0 are mapped into triangles, thus Th,t is a valid triangulation,
under the requirement of invertibility of the map (which is assured if the domain is
convex and the wall displacements are small).
As for the time evolution, we may adopt a linear time variation within each time slab
[t k , t k+1 ] by setting
t − tk t − t k+1
At = At k+1 − At k , t ∈ t k , t k+1 .
t t
Then, the corresponding domain velocity wh will be constant on each time slab.
We are now in the position of describing a possible finite element scheme for both
the structure and the fluid problem, to be adopted in the sub-structuring algorithm. We
first give more details on the adopted finite element discretisation.
19.1. The discretisation of the structure
For the structure we consider a mid-point scheme. We introduce the additional variable
η̇k which is the approximation of the displacement velocity at time t k .
The time advancing scheme reads:
∀k 0 find ηk+1 and η̇k+1 that satisfy the following system:
ηk+1 − ηk η̇k + η̇k+1

= , (19.2a)
t 2
η̇k+1 − η̇k ∂ 2 ηk + ηk+1 ηk+1 + ηk ∂ 2 η̇k+1 + η̇k
−a 2 +b −c 2
t ∂z 2 2 ∂z 2
=H k+1/2
, (19.2b)
with

ηk+1 |z=0 = α t k+1 , ηk+1 |z=L = β t k+1 , (19.3a)
S ECTION 19 83
and
∂ k+1 ∂ k+1
η̇k+1 |z=0 = α t , η̇k+1 |z=L = β t , (19.3b)
∂t ∂t
while the value of η0 and η̇0 are given by the initial conditions.
Here, H k+1/2 is a suitable approximation of H at time t k + 12 t which in the context
of a sub-structuring iteration for the coupled problem is a known quantity and whose
calculation from the Navier–Stokes data will be made precise later.
System (19.2) is then discretised in space by taking ηhk ∈ Sh and η̇hk ∈ Sh . We set
Sh = {sh ∈ Sh : sh (0) = 0, sh (L) = 0} and the finite element problem reads:
0
For all k 0 find ηhk+1 ∈ Sh and η̇hk+1 ∈ Sh that satisfy the following system:
k+1
2ηh − t η̇hk+1 , sh = 2ηhk + t η̇hk , sh , (19.4a)

1 k+1 b k+1 1 ∂ηhk+1 ∂ η̇k+1 ∂sh
η̇h + ηh , sh + a +c h ,
t 2 2 ∂z ∂z ∂z

1 k b k 1 ∂ηh k
∂ η̇k ∂sh
= H k+1/2, sh + η̇h + ηh , sh − a +c h, , (19.4b)
t 2 2 ∂z ∂z ∂z
∀sh ∈ Sh0 , together with the boundary conditions

ηhk+1 |z=0 = α t k+1 , ηhk+1 |z=L = β t k+1 , (19.5a)
∂ ∂
η̇hk+1 |z=0 = α t k+1 , η̇hk+1 |z=L = β t k+1 , (19.5b)
∂t ∂t
and the initial conditions
ηh0 = πSh η0 , ηh0 = πSh η0 ,

being π Sh the standard interpolation operator upon Sh .
19.2. The discretisation of the fluid problem
In the frame of our splitting scheme the velocity field at Γtw as well as the current do-
main configuration are provided by the calculation of ηh ; they can thus be considered
as given data. We consider the following finite element spaces. Q h is the space of con-

tinuous piece-wise linear finite elements, while Vh is that of vector functions whose
components are in the space V h of continuous piece-wise quadratic (or P1-isoP2) finite
f
elements. Both refer to the triangulation T0,h of Ω0 . For a precise definition of these fi-
nite element spaces the reader may refer to Q UARTERONI and VALLI [1994] or B REZZI
and F ORTIN [1991].
We will also need to define
V0h = {ṽh ∈
Vh : ṽh |∂Ω0 \Γ in = 0}
0
w
and the space Γ0
V h .
formed by function in Γ0w which are the trace of a function in V
h
The corresponding spaces on the current configuration will be given by

h },
Qh,t = {qh : qh ◦ At ∈ Q Vh,t = {vh : vh ◦ At ∈
Vh },
and analogously for V0h,t .
R EMARK 19.2. The functions belonging to Qh,t and Vh,t depend also on time through
the ALE mapping. A thorough presentation of finite element spaces in an ALE frame-
work is contained in F ORMAGGIA and N OBILE [1999] and N OBILE [2001].
We will employ an implicit Euler time advancing scheme with a semi-explicit treat-
ment of the convective term. Let us assume that the solution (ukh , phk ) at time step t k is
known, as well as the domain configuration Ωt k+1 ,h at time t k+1 (and thus the corre-
sponding ALE map).
The numerical solution at t k+1 can be computed as follows:
Find uk+1
h ∈ Vh,t k+1 and phk+1 ∈ Qh,t k+1 such that
1 k+1
u , ṽh k+1 − ck+1/2 wk+1/2 , uk+1 , ṽh + ck+1 uk , uk+1 , ṽh
t

+ dk+1/2 wk+1/2 , uk+1 , ṽh + bk+1 ṽh , pk+1 + ak+1 uk+1 , ṽh
1 k
= f k+1 , ṽh k+1 + u , ṽh k , ∀ṽh ∈
V0h (19.6a)
t

bk+1 uk+1 , q̃h = 0, ∀q̃h ∈ Q h , (19.6b)
and
uk+1
h = gk+1 in
h , on Γt k+1 , (19.7a)
Γ w
uk+1
h = Πh 0 η̇hk+1 ◦ A−1 e ,
t k+1 r
on Γtwk+1 . (19.7b)
We have defined

(w, ṽ)k = w · ṽ ◦ A−1
tk
,
Ωt k

ck (w, z, ṽ) = (w·∇)z · ṽ ◦ A−1
tk
,
Ωt k

dk (w, z, ṽ) = (div w)z · ṽ ◦ A−1
tk
,
Ωt k

bk (w, q̃) = div w q̃ ◦ A−1
tk
, w, q) =
bk ( w ◦ A−1
div tk
q,
Ωt k Ωt k

ak (w, ṽ) = 2νD(w) : D ṽ ◦ A−1
tk
.
Ωt k
The function gk+1

his the finite element interpolant of the boundary data g(t k+1 ) on
Γ0w w
the space of restrictions of Vh,t k+1 on Γtin
k+1 . Moreover, Πh
Γ0 is the interpo-
: Sh → V h
lation operator required to project the discrete vessel velocity computed by the structure
S ECTION 19 85
solver on the trace space of discrete fluid velocity on the vessel wall. Since we are using
geometrically conforming finite elements, this operator is quite simple to build up.
It is understood that when the approximation of u and w in (19.6) are not evaluated at
the same time as the integral, they need to be mapped on the correct domain by means
of the ALE transformation.
R EMARK 19.3. The term involving the domain velocity w has been computed on the
intermediate geometry Ωt k+1/2 in order to satisfy the so-called Geometry Conservation
Law (GCL) (G UILLARD and FARHAT [2000]). A discussion on the significance of the
GCL for the problem at hand may be found in N OBILE [2001].
19.3. Recovering the forcing term for the vessel wall
We need now to compute the forcing term H k+1/2 in (19.4) as the residual of the discrete
momentum equation (19.6b) for time step t k+1 . Let us define
1 k 1 k+1
Rhk+1 (ṽh ) = f k+1 , ṽh k+1 + u , ṽh k − u , ṽh k+1
t t
k+1/2 k+1 k k+1
+ ck+1/2 w , u , ṽh − ck+1 u , u , ṽh
k+1/2 k+1
− dk+1/2 w , u , ṽh − bk+1 ṽh , pk+1 − ak+1 uk+1 , ṽh , ∀ṽh ∈
Vh .
Note that Rhk+1 (ṽh ) = 0, for all ṽh ∈

V0h . We define the following operator:
Γ w
S h : Sh →
Vh , S h sh = Rh Πh 0 sh er ,
Γw
0 →V
where Rh : V h is a finite element extension operator such that
h
Γw
(Rh vh )|Γ0w = vh , 0 ,
∀vh ∈ V h
for instance the one obtained by extending by zero at all internal nodes (see Q UAR -
TERONI and VALLI [1999]). We then take
k+1/2 ω k+1
H , sh = R (S h sh ) + Rhk (S h sh ) . (19.8)
2 h
19.4. The algorithm
We are now in the position of describing an iterative algorithm for the solution of the
coupled problem. As usual, we assume to have all quantities available at t = t k , k 0,
provided either by previous calculations or by the initial data and we wish to advance
to the new time step t k+1 . For ease of notation we here omit the subscript h, with the
understanding that we are referring exclusively to finite element quantities.
The algorithm requires to choose a tolerance τ > 0, which is used to test the con-
vergence of the procedure, and a relaxation parameter 0 < θ 1. In the following, the
subscript j 0 denotes the sub-iteration counter.
The algorithm reads:
(A1) Extrapolate the vessel wall structure displacement and velocity:

k+1 k+1
η(0) = ηk + t η̇k , η̇(0) = η̇k .
(A2) Set j = 0.
k+1
(A2.1) By using η(j ) , compute the new grid for the fluid domain Ωt and the
ALE map by solving (19.1).
(A2.2) Solve the Navier–Stokes problem (19.6) to compute uk+1 k+1
(j +1) and p(j +1) ,
k+1
using as velocity on the wall boundary the one calculated from η̇(j ) .
k+1 k+1
(A2.3) Solve (19.4) to compute η∗ and η̇∗ using as forcing term the one
recovered from uk+1 k+1
(j +1) and p(j +1) using (19.8).
k+1 k+1 − η̇k+1
(A2.4) Unless η∗k+1 − η(j ) L2 (Γ w ) + η̇∗ (j ) L2 (Γ w ) τ , set
0 0
+1) = θ η(j ) + (1 − θ )η∗ ,

k+1 k+1 k+1
η(j
k+1 k+1
+1) = θ η̇(j ) + (1 − θ )η̇∗ ,
k+1
η̇(j
and j ← j + 1. Then return to step (A2.1).
(A3) Set
ηk+1 = η∗k+1 , η̇k+1 = η̇∗k+1 .
uk+1 = uk+1
(j +1) ,
k+1
pk+1 = p(j +1) .
If the algorithm converges, limj →∞ uk+1(j ) = u

k+1 and lim k+1
j →∞ η(j ) = η
k+1 , where
uk+1 and ηk+1 are the solution at time step t k+1 of the coupled problem.
The algorithm entails, at each sub-iteration, the computation of the generalised string
equation (19.4)–(19.5), the Navier–Stokes equations and the solution of two Laplace
equations (19.1), one for every displacement component.
Improvements on the computational efficiency of the coupled procedure just de-
scribed may be obtained either by employing standard acceleration techniques like
Aitken extrapolation, or by using an altogether different approach to the non-linear
problem (like Newton–Krylov techniques or multilevel schemes). The matter is still
the subject of current active research investigations.
More explicit schemes for the fluid–structure interaction problem, known as “ser-
ial staggered” procedures, have been successfully applied to aeroelastic 2D and 3D
problems (FARHAT, L ESOINNE and M AMAN [1995], FARHAT and L ESOINNE [2000],
P IPERNO and FARHAT [2001]). However, it has been found that in the case of an incom-
pressible fluid they become unstable when the density of the structure mass is compa-
rable to that of the fluid (L E TALLEC and M OURO [2001]), which is unfortunately our
situation. An analysis of decoupling technique for unsteady fluid structure interaction,
carried out on a simplified, yet representative, one-dimensional model may be found in
G RANDMONT, G UIMET and M ADAY [2001].
C HAPTER VI
20. One-dimensional models of blood flow in arteries
In this section we introduce a simple 1D model to describe the flow motion in arter-
ies and its interaction with the wall displacement. In the absence of branching, a short
section of an artery may be considered as a cylindrical compliant tube. As before we
denote by I = (t0 , t1 ) the time interval of interest and by Ωt the spatial domain which
is supposed to be a circular cylinder filled with blood. The reason why one-dimensional
models for blood flow may be attractive is that full 3D investigations are quite compu-
tationally expensive. Yet, in many situations we might desire to have just information of
the evolution of averaged quantities along the arterial tree, such as mass flux and aver-
age pressure. In this context simplified models are able to provide an reasonable answer
in short times.
As already done in Section 15, we will employ cylindrical coordinates and indicate
with er , eθ and ez the radial, circumferential and axial unit vectors, respectively, and
with (r, θ, z) the corresponding coordinates system. The vessel extends from z = 0 to
z = L and the vessel length L is constant with time.
The basic model is deduced by making the following assumptions, some of which
are analogous to the ones made in Section 16:
(A1) Axial symmetry. All quantities are independent from the angular coordinate θ .
As a consequence, every axial section z = const. remains circular during the
wall motion. The tube radius R is a function of z and t.
(A2) Radial displacements. The wall displaces along the radial direction solely, thus
at each point on the tube surface we may write η = ηer , where η = R − R0 is
the displacement with respect to the reference radius R0 .
(A3) Constant pressure. We assume that the pressure P is constant on each section,
so that it depends only on z and t.
(A4) No body forces. We neglect body forces (the inclusion of the gravity force,
if needed, is straightforward); thus we put f = 0 in the momentum equation
(11.1a).
(A5) Dominance of axial velocity. The velocity components orthogonal to the z axis
are negligible compared to the component along z. The latter is indicated by uz
and its expression in cylindrical coordinates reads

r
uz (t, r, z) = ū(t, z)s , (20.1)
R(t, z)
87
88 A. Quarteroni and L. Formaggia C HAPTER VI
where ū is the mean velocity on each axial section and s : R → R is a velocity

profile.6
A generic axial section will be indicated by S = S(t, z). Its measure A is given by
2
A(t, z) = meas S(t, z) = πR 2 (t, z) = π R0 (z) + η(t, z) . (20.2)
The mean velocity ū is then given by

−1
ū = A uz dσ,
S
and from (20.1) and the definition of ū it follows that
1
1
s(y)y dy = .
0 2
We will indicate with α the momentum-flux correction coefficient (sometimes called
Coriolis coefficient), defined as
2 2
S uz dσ s dσ
α= = S , (20.3)
Aū2 A
where the dependence of the various quantities on the spatial and time coordinates is
understood. It is immediate to verify that α 1. In general, this coefficient will vary in
time and space, yet in our model it is taken constant as a consequence of (20.1).
One possible choice for the profile law is the parabolic profile s(y) = 2(1 − y 2 ),
which corresponds to the Poiseuille solution characteristic of steady flows in circular
tubes. In this case we have α = 4/3. However, for blood flow in arteries it has been
found that the velocity profile is, on average, rather flat. Indeed, a profile law often used
for blood flow in arteries (see, for instance, S MITH , P ULLAN and H UNTER [2003]) is a
power law of the type s(y) = γ −1 (γ + 2)(1 − y γ ), with typically γ = 9 (the value γ = 2
gives again the parabolic profile). Correspondingly, we have α = 1.1. Furthermore, we
will see that the choice α = 1, which indicates a completely flat velocity profile, would
lead to a certain simplification in our analysis.
The mean flux Q, defined as

Q = uz dσ = Aū,
S
is one of the main variables of our problem, together with A and the pressure P .
20.1. The derivation of the model
There are (at least) three ways of deriving our model. The first one moves from the in-
compressible Navier–Stokes equations with constant viscosity and performs an asymp-
totic analysis by assuming that the ratio R0 /L is small, thus discarding the higher order
terms with respect to R0 /L (see BARNARD , H UNT, T IMLAKE and VARLEY [1966]).
6 The fact that the velocity profile does not vary is in contrast with experimental observations and numerical
results carried out with full scale models. However, it is a necessary assumption for the derivation of the
reduced model. One may then think s as being a profile representative of an average flow configuration.
S ECTION 20 89
F IG . 20.1. A longitudinal section (θ = const.) of the tube and the portion between z = z∗ − dz/2 and
z = z∗ + dz/2 used for the derivation of the 1D reduced model.
The second approach derives the model directly from the basic conservation laws writ-
ten in integral form. The third approach consists of integrating the Navier–Stokes equa-
tions on a generic section S.
We will indicate with Γtw the wall boundary of Ωt , which now reads

Γtw = (r, θ, z): r = R(z, t), θ ∈ [0, 2π), z ∈ (0, L)
while n is the outwardly oriented normal to ∂Ωt . Under the previous assumption, the
momentum and continuity equations along z are:
∂uz 1 ∂P
+ div(uz u) + − ν uz = 0, z ∈ (0, L), t ∈ I, (20.4a)
∂t ρ ∂z
div u = 0, z ∈ (0, L), t ∈ I, (20.4b)
and on the tube wall we have
u = η̇, on Γtw , t ∈ I.
We have written the convective term in divergence form, like in (8.12), because it sim-
plifies the further derivation.
To ease notation, in this section we will omit to explicitly indicate the time depen-
dence, with the understanding that all variables are considered at time t. Let us con-
sider the portion P of Ωt , sketched in Fig. 20.1, comprised between z = z∗ − dz/2 and
z = z∗ + dz/2, with z∗ ∈ (0, L) and dz > 0 small enough so that z∗ + dz/2 < L and
z∗ − dz/2 > 0. The part of ∂P laying on the tube wall is indicated by ΓPw . The reduced
model is derived by integrating (20.4b) and (20.4a) on P and passing to the limit as
dz → 0, assuming that all quantities are smooth enough.
We will first illustrate a result derived from the application of the ALE transport
theorem (Theorem 18.1) to P.
L EMMA 20.1. Let f : Ωt × I → R be an axisymmetric function, i.e., ∂f /∂θ = 0. Let

us indicate by fw the value of f on the wall boundary and by f¯ its mean value on each
axial section, defined by

¯
f =A −1
f dσ .
S
We have the following relation:
∂ ∂f
(Af¯) = A + 2πR η̇fw . (20.5)
∂t ∂t
In particular, taking f = 1 yields
∂A
= 2πR η̇. (20.6)
∂t
P ROOF. The application of (18.10) to P gives

d ∂f
f= + f g · n, (20.7)
dt P P ∂t ∂P
where g denotes the velocity of the boundary of P, i.e.,

η̇ on ΓPw ,
g= (20.8)
0 on ∂P \ ΓPw .
Then, by applying the mean-value theorem to both sides of (20.7), we have

d ∂f
A(z∗ )f¯(z∗ ) dz + o(dz) = A + o(dz) + f η̇er · n.
dt ∂t ΓPw
We recall relation (16.5), already used in the derivation of the models for the wall struc-
ture dynamics, to write

f η̇er · n
ΓPw
z∗ +dz/2
2π
= f η̇R dz dθ = 2π η̇(z∗ )R(z∗ )fw (z∗ ) dz + o(dz) . (20.9)
0 z∗ −dz/2
By substituting into (20.7), dividing by dz and passing to the limit as dz → 0, we obtain
the desired result.
We are now ready to derive our reduced model. We start first from the continuity
equation. Using the divergence theorem, we obtain

0= div u = − uz + uz + u ·n
P S− S+ ΓPw

=− uz + uz + η̇ · n. (20.10)
S− S+ ΓPw
We have exploited (20.8) and the fact that n = −ez on S − while n = ez on S + . Now,
since η̇ = η̇er , we deduce

∂
η̇ · n = 2η̇πR(z∗ ) dz + o(dz) = (by (20.6)) = A(z∗ ) dz + o(dz).
ΓPw ∂t
S ECTION 20 91
By substituting into (20.10), using the definition of Q, and passing to the limit as
dz → 0, we finally obtain
∂A ∂Q
+ = 0,
∂t ∂z
which is the reduced form of the continuity equation.
We will now consider all terms in the momentum equation in turn. Again, we will
integrate them over P and consider the limit as dz tends to zero,

∂uz d d
= uz − uz g · n = uz .
P ∂t dt P ∂P dt P
In order to eliminate the boundary integral, we have exploited the fact that uz = 0 on
ΓPw and g = 0 on S − and S + . We may then write

∂uz ∂ ∂Q ∗
= A(z∗ )ū(z∗ ) dz + o(dz) = (z ) dz + o(dz).
P ∂t ∂t ∂t
Moreover, we have

div(uz u) = uz u · n = − u2z + u2z + uz g · n
P ∂P S− S+ ΓPw

∗ dz 2 ∗ dz ∗ dz 2 ∗ dz
=α A z + ū z + −A z − ū z −
2 2 2 2
∂αAū2 ∗
= (z ) dz + o(dz).
∂z
Again, we have exploited the condition uz = 0 on ΓPw .
Since the pressure is assumed to be constant on each section, we obtain

∂P
=− P+ P+ P nz
P ∂z S− S+ ΓPw

∗ dz ∗ dz ∗ dz ∗ dz
=A z + P z + −A z − P z −
2 2 2 2

+ P nz . (20.11)
ΓPw
Since

nz = 0,
∂P
we may write that

∗ ∗
P nz = P (z ) nz + o(dz) = −P (z ) nz + o(dz)
ΓPw ΓPw ∂ P \ΓPw

dz dz
= −P (z∗ ) A z∗ + − A z∗ − + o(dz).
2 2
By substituting the last result into (20.11), we have

∂P dz dz dz dz
= A z∗ + P z∗ + − A z∗ − P z∗ −
P ∂z 2 2 2 2

dz dz
− P (z∗ ) A z∗ + − A z∗ − + o(dz)
2 2
∂(AP ) ∗ ∂A ∗ ∂P ∗
= (z ) dz − P (z∗ ) (z ) dz + o(dz) = A (z ) dz + o(dz).
∂z ∂z ∂z
We finally consider the viscous term,

∂uz ∂uz
uz = ∇uz · n = − + + ∇uz · n.
P ∂P S − ∂z S + ∂z w
ΓP
We neglect ∂uz /∂z by assuming that its variation along z is small compared to the
other terms. Moreover, we split n into two vector components, the radial component
nr = nr er and nz = n − nr . Owing to the cylindrical geometry, n has no component
along the circumferential coordinate and, consequently, nz is indeed oriented along z.
We may thus write

uz = (∇uz · nz + ∇uz · er nr ) dσ.
P ΓPw
Again, we neglect the term ∇uz ·nz , which is proportional to ∂uz /∂z. We recall relation
(20.1) to write
z∗ +dz/2
uz = nr ∇uz · er dσ = ūR −1 s (1)n · er dσ = 2π ūs (1) dz,
P ΓPw ΓPw z∗ −dz/2
where we have used the relation nr dσ = 2πR dz and indicated by s the first derivative
of s.
Then,

uz ≈ 2π ū(z∗ )s (1) dz.
P
By substituting all results into (20.4a), dividing all terms by dz and passing to the limit
as dz → 0, we may finally write the momentum equation of our one-dimensional model
as follows:
∂Q ∂(αAū2 ) A ∂P
+ + + Kr ū = 0,
∂t ∂z ρ ∂z
where
Kr = −2πνs (1)
is a friction parameter, which depends on the type of profile chosen, i.e., on the choice
of the function s in (20.1). For a profile law given by s(y) = γ −1 (γ + 2)(1 − y γ ), we
have Kr = 2πν(γ + 2). In particular, for a parabolic profile Kr = 8πν, while for γ = 9
we obtain Kr = 22πν.
S ECTION 20 93
To conclude, the final system of equations reads

∂A ∂Q
+ = 0, z ∈ (0, L), t ∈ I, (20.12a)
∂t ∂z

∂Q ∂ Q2 A ∂P Q
+α + + Kr = 0, z ∈ (0, L), t ∈ I, (20.12b)
∂t ∂z A ρ ∂z A
where the unknowns are A, Q and P and α is here taken constant.
20.2. Accounting for the vessel wall displacement
In order to close system (20.12), we provide a relation for the pressure. A possibility is to
resort to an algebraic relation linking pressure to the wall deformation and consequently
to the vessel section A.
More generally, we may assume that the pressure satisfies a relation like

P (t, z) − Pext = ψ A(t, z); A0(z), β(z) , (20.13)
where we have outlined that the pressure will in general depend also on A0 = πR02
and on a set of coefficients β = (β0 , β1 , . . . , βp ), related to physical and mechanical
properties, that are, in general, given functions of z. Here Pext indicates, as in Section 15,
the external pressure. We require that ψ be (at least) a C 1 function of all its arguments
and be defined for all A > 0 and A0 > 0, while the range of variation of β will depend
by the particular mechanical model chosen for the vessel wall.
Furthermore, we require that for all allowable values of A, A0 and β,
∂ψ
> 0 and ψ(A0 ; A0, β) = 0. (20.14)
∂A
By exploiting the linear elastic law provided in (16.16), with the additional simplify-
ing assumption (16.15), and using the fact that
√ √
η = A − A0 / π, (20.15)
we can obtain the following expression for ψ:
√ √
A − A0
ψ(A; A0 , β0 ) = β0 . (20.16)
A0
We have identified β with the single parameter
√
πh0 E
β0 = .
1 − ξ2
The latter depends on z only in those cases where the Young modulus E or the vessels
thickness h0 are not constant.
For ease of notation, the dependence of A, A0 and β from their arguments will be
understood. It is immediate to verify that all the requirements in (20.14) are indeed
satisfied.
Another commonly used expression for the pressure-area relationship is given
by H AYASHI , H ANDA , NAGASAWA and O KUMURA [1980], S MITH , P ULLAN and
H UNTER [2003]:
β1
A
ψ(A; A0 , β) = β0 −1 .
A0
In this case, β = (β0 , β1 ), where β0 > 0 is an elastic coefficient while β1 > 0 is normally
obtained by fitting the stress-strain response curves obtained by experiments.
Another alternative formulation (L ANGEWOUTERS , W ESSELING and G OEDHARD
[1984]) is

A − A0
ψ(A; A0 , β) = β0 tan π ,
2A0
where again the coefficients vector β reduces to the single coefficient β0 .
In the following, whenever not strictly necessary we will omit to indicate the depen-
dence of the various quantities on A0 and β, which is however always understood.
20.3. The final model
By exploiting relation, (20.12), we may eliminate the pressure P from the momentum
equation. To that purpose we will indicate by c1 = c1 (A; A0, β) the following quantity:

A ∂ψ
c1 = , (20.17)
ρ ∂A
which has the dimension of a velocity and, as we will see later on, is related to the speed
of propagation of simple waves along the tube.
By simple manipulations (20.12) may be written in quasi-linear form as follows:
∂ ∂U
U + H(U ) + B(U ) = 0, z ∈ (0, L), t ∈ I (20.18)
∂t ∂z
where,

A
U= ,
Q
 
0 1 0 1
2
H(U ) = A ∂ψ = 2 Q Q , (20.19)
− α ū2 2α ū c1 − α 2α
ρ ∂A A A
and

0
B(U ) = Q A ∂ψ dA0 A ∂ψ dβ .
KR + +
A ρ ∂A0 dz ρ ∂β dz
Clearly, if A0 and β are constant the expression for B becomes simpler. A conservation
form for (20.18) may be found as well and reads
∂U ∂
+ F (U ) + S(U ) = 0, (20.20)
∂t ∂z
S ECTION 20 95
where
Q
F (U ) = Q2
+ C1
α
A
is the vector of fluxes,

0
S(U ) = B(U ) − ∂C1 dA0 ∂C1 dβ
+ ,
∂A0 dz ∂β dz
and C1 is a primitive of c12 with respect to A, given by
A
C1 (A; A0, β) = c12 (τ ; A0, β) dτ.
A0
Again, if A0 and β are constant, the source term S simplifies and becomes S = B.
System (20.20) allows to identify the vector U as the conservation variables of our
problem.
R EMARK 20.1. In the case we use relation (20.16), we have

β0 β0
c1 = A1/4, C1 = A3/2. (20.21)
2ρA0 3ρA0
L EMMA 20.2. If A 0, the matrix H possesses two real eigenvalues. Furthermore, if

A > 0 the two eigenvalues are distinct and (20.18) is a strictly hyperbolic system of
partial differential equations.
P ROOF. By straightforward computations, we have the following expression for the

eigenvalues of H:
λ1,2 = α ū ± cα , (20.22)
where
%
cα = c12 + ū2 α(α − 1).
Since α 1, cα is a real number. If cα > 0 the two eigenvalues are distinct. A sufficient
condition to have cα > 0 is c1 > 0 and, thanks to the definition of c1 and (20.14), this is
always true if A > 0. If α = 1, this condition is also necessary.
The existence of a complete set of (right and left) eigenvectors is an immediate con-
sequence of H having distinct eigenvalues.
R EMARK 20.2. System (20.12) shares many analogies with the 1D compressible Euler
equations, after identifying the section area A with the density. The equivalence is not
complete since the term ∂P /∂z in the Euler equations is here replaced by A∂P /∂z.
20.3.1. Characteristics analysis

Let (l1 , l2 ) and (r1 , r2 ) be two couples of left and right eigenvectors of the matrix H in
(20.19), respectively. The matrices L, R and are defined as
T
l1 λ1 0
L= T , R = [ r1 r2 ] , = diag(λ1 , λ2 ) = . (20.23)
l2 0 λ2
Since right and left eigenvectors are mutually orthogonal, without loss of generality we
choose them so that LR = I. Matrix H may then be decomposed as
H = R L, (20.24)
and system (20.18) written in the equivalent form
∂U ∂U
L + L + LB(U ) = 0, z ∈ (0, L), t ∈ I. (20.25)
∂t ∂z
If there exist two quantities W1 and W2 which satisfy
∂W1 ∂W2
= l1 , = l2 , (20.26)
∂U ∂U
we will call them characteristic variables of our hyperbolic system. We point out that
in the case where the coefficients A0 and β are not constant, W1 and W2 are not au-
tonomous functions of U .
By setting W = [W1 , W2 ]T , system (20.25) may be elaborated into
∂W ∂W
+ + G = 0, z ∈ (0, L), t ∈ I, (20.27)
∂t ∂z
where
∂W dA0 ∂W dβ
G = LB − − . (20.28)
∂A0 dz ∂β dz
We note that the extra terms on the right-hand side are a consequence of the fact that
the characteristic variables depend parametrically on the coefficient A0 and β, which
may by a function of z, and thus
∂W ∂W ∂U ∂W dA0 ∂W dβ ∂U ∂W dA0 ∂W dβ
= + + =L + + .
∂z ∂U ∂z ∂A0 dz ∂β dz ∂z ∂A0 dz ∂β dz
In the case where B = 0 and the coefficients A0 and β are constant, (20.27) takes the
simpler form
∂W ∂W
+ = 0, z ∈ (0, L), t ∈ I, (20.29)
∂t ∂z
which component-wise reads
∂Wi ∂Wi
+ λi = 0, z ∈ (0, L), t ∈ I, i = 1, 2. (20.30)
∂t ∂z
S ECTION 20 97
R EMARK 20.3. From definition (20.26) and the fact that the left and right eigenvectors
li and ri are mutually orthogonal it follows that
∂W1
(U ) · r2 (U ) = 0,
∂U
thus W1 is a 2-Riemann invariant of our hyperbolic system (G ODLEWSKI and R AVIART
[1996]). Analogously, one may show that W2 is a 1-Riemann invariant.
From (20.30) we have that W1 and W2 are constant along the two characteristic
curves in the (z, t) plane described by the differential equations
dz dz
= λ1 and = λ2 ,
dt dt
respectively. In the more general case (20.27) we may easily show that W1 and W2
satisfy a coupled system of ordinary differential equations.
The expression for the left eigenvectors l1 and l2 is given by

c − α ū −cα − α ū
l1 = ζ α , l2 = ζ ,
1 1
where ζ = ζ (A, ū) is any arbitrary smooth function of its arguments with ζ > 0. Here
we have expressed l1 and l2 as functions of (A, ū) instead of (A, Q) as is more conve-
nient for the next developments. Thus, relations (20.26) become
∂W1 ∂W1
= ζ cα − ū(α − 1) , = ζ A, (20.31a)
∂A ∂ ū
∂W2 ∂W2
= ζ −cα − ū(α − 1) , = ζ A. (20.31b)
∂A ∂ ū
For a hyperbolic system of two equations it is always possible to find the character-
istic variables (or, equivalently, the Riemann invariants) locally, that is in a sufficiently
small neighbourhood of any point U (G ODLEWSKI and R AVIART [1996], L AX [1973]),
yet the existence of global characteristic variables is not in general guaranteed. How-
ever, in the special case α = 1, (20.31) takes the much simpler form
∂W1 ∂W1
= ζ c1 , = ζ A,
∂A ∂ ū
∂W2 ∂W2
= −ζ c1 , = ζ A.
∂A ∂ ū
Let us show that a set of global characteristic variables for our problem does exist in
this case. We remind that the characteristic variable W1 exists if and only if
∂ 2 W1 ∂ 2 W1
= ,
∂A∂ ū ∂ ū∂A
for all allowable values of A and ū. Since now c1 does not depend on ū, the above
condition yields
∂ζ ∂ζ
c1 =ζ +A .
∂ ū ∂A
In order to satisfy this relation, it is sufficient to take ζ = ζ(A) such that ζ =

−A(∂ζ /∂A). A possible instance is ζ = A−1 . The resulting differential form is
c1
∂W1 = ∂A + ∂ ū,
A
and by proceeding in the same way for W2 , we have
c1
∂W2 = − ∂A + ∂ ū,
A
To integrate it in the (A, ū) plane, we need to fix the value at a reference state, for
instance W1 = W2 = 0 for (A, ū) = (A0 , 0). We finally obtain
A A
c1 (τ ) c1 (τ )
W1 = ū + dτ, W2 = ū − dτ. (20.32)
A0 τ A0 τ
R EMARK 20.4. If we adopt relation (20.16) and use the expression for c1 given in
(20.21), after simple computations we have
W1 = ū + 4(c1 − c1,0 ), W2 = ū − 4(c1 − c1,0 ), (20.33)
where c1,0 is the value of c1 corresponding to the reference vessel area A0 .
Under physiological conditions, typical values of the flow velocity and mechanical
characteristics of the vessel wall are such that cα α ū. Consequently, λ1 > 0 and
λ2 < 0, i.e., the flow is sub-critical everywhere. Furthermore, the flow is smooth. Dis-
continuities, which would normally appear when treating a non-linear hyperbolic sys-
tem, do not have indeed the time to form in our context because of the pulsatility of
the boundary conditions. It may be shown (C ANIC and K IM [2003]) that, for the typi-
cal values of the mechanical and geometric parameters in physiological conditions and
the typical vessel lengths in the arterial tree, the solution of our hyperbolic system re-
mains smooth, in accordance to what happens in the actual physical problem (which is
however dissipative, a feature which has been neglected in our one-dimensional model).
In the light of the previous considerations, from now on we will always assume sub-
critical regime and smooth solutions.
20.3.2. Boundary conditions

System (20.12) must be supplemented by proper boundary conditions. The number of
conditions to apply at each end equals the number of characteristics entering the domain
through that boundary. Since we are only considering sub-critical flows, we need to
impose exactly one boundary condition at both z = 0 and z = L.
An important class of boundary conditions, called non-reflecting or ‘absorbing’, are
those that allow the simple wave associated to the outgoing characteristic to exit the
computational domain with no reflections. Following T HOMPSON [1987], H EDSTROM
[1979], non-reflecting boundary conditions for one-dimensional systems of non-linear
hyperbolic equation like (20.20) may be written as

∂U ∂U
l1 + S(U ) = 0 at z = 0, l2 + S(U ) = 0 at z = L,
∂t ∂t
S ECTION 20 99
for all t ∈ I . When S = 0 these conditions are equivalent to impose a constant value
(typically set to zero) to the incoming characteristic variable. When S = 0 they take into
account the “natural variation” of the characteristic variables due to the presence of the
source term. A boundary condition of this type is quite convenient at the outlet section.
At the inlet instead one usually desires to impose values of pressure or mass flux
derived from measurements or other means. Let us suppose that z = 0 is an inlet section
(the following discussion may be readily extended to the boundary z = L). Whenever an
explicit formulation of the characteristic variables is available, the boundary condition
may be expressed directly in terms of the entering characteristic variable W1 , i.e., for
all t ∈ I ,
W1 (t) = g1 (t) at z = 0, (20.34)
g1 being a given function. However, seldom one has directly g1 at disposal, as the
available boundary data is normally given in terms of physical variables. Let us suppose
that we know the time variation of both pressure and mass flux at that boundary (for
instance, taken from measurements). We may derive the corresponding value of g1 using
directly the definition of the characteristic variable W1 . If Pm = Pm (t) and Qm = Qm (t)
are the measured average pressure and mass flux at z = 0 for t ∈ I and W1 (A, Q)
indicates the characteristic variable W1 as function of A and Q, we may pose

g1 (t) = W1 ψ −1 Pm (t) − Pext , Qm (t) , t ∈ I,
in (20.34). This means that Pm and Qm are not imposed exactly at z = 0 (this would
not be possible since our system accounts for only one boundary condition at each end
of the computational domain), yet we require that at all times t the value of A and Q at
z = 0 lies on the curve in the (A, Q) plane defined by

W1 (A, Q) − W1 ψ −1 Pm (t) − Pext , Qm (t) = 0.
If instead one has at disposal the time history q(t) of a just one physical variable
φ = φ(A, Q), the boundary condition

φ A(t), Q(t) = q(t), ∀t ∈ I, at z = 0,
is admissible under certain restrictions (Q UARTERONI and VALLI [1994]), which in our
case reduce to exclude the case where φ may be expressed as function of only W2 . In
particular, it may be found that for the problem at hand the imposition of either average
pressure or mass flux are both admissible.
R EMARK 20.5. If the integration of (20.26) is not feasible (as, for instance, in the case
α = 1), one may resort to the pseudo-characteristic variables (Q UARTERONI and VALLI
[1994]), Z = [Z1 , Z2 ]T , defined by linearising (20.26) around an appropriately chosen
reference state. One obtains
Z = Z + L(U )(U − U ), (20.35)
where U is the chosen reference state and Z the corresponding value for Z. One may
then use the pseudo-characteristic variables instead of W , by imposing
Z1 (t) = g1 (t) at z = 0, Z2 (t) = 0 at z = L.
In the context of a time advancing scheme for the numerical solution of (20.20) the
pseudo-characteristics are normally computed linearising around the solution computed
at the previous time step.
R EMARK 20.6. When considering the numerical discretisation, we need in general to

provide an additional equation at each end point in order to close the resulting alge-
braic system. Typically, this extra relation is provided by the so-called compatibility
conditions (Q UARTERONI and VALLI [1994]), which read as follows:

T ∂ ∂U
l2 U +H + B = 0, z = 0, t ∈ I, (20.36a)
∂t ∂z

∂ ∂U
lT1 U +H + B = 0, z = L, t ∈ I. (20.36b)
∂t ∂z
20.3.3. Energy conservation for the 1D model

Most of the results presented in this section are taken from F ORMAGGIA , G ERBEAU ,
N OBILE and Q UARTERONI [2001] and C ANIC and K IM [2003].
L EMMA 20.3. Let us consider the hyperbolic problem (20.18) and assume that the
initial and boundary conditions are such that ∀z ∈ (0, L),
A(0, z) > 0 and A(t, 0) > 0, A(t, L) > 0, ∀t ∈ I,

and that the solution U is smooth for all (t, z) ∈ I × (0, L). Then A(t, z) > 0 for all
(t, z) ∈ I × (0, L).
P ROOF. Let us suppose that we have A(t ∗ , z∗ ) = 0 at a generic point (t ∗ , z∗ ) ∈ I ×

(0, L). From the definition of λ1 and λ2 the line l = {(t, z): z = zu (t)} satisfying
dzu
(t) = ū t, zu (t)
dt
and ending at the point (t ∗ , z∗ ), lies between the two characteristic curves passing
through the same point. Therefore, it completely lies inside the domain of dependence of
(t ∗ , z∗ ) and either intersects the segment z ∈ (0, L) at t = 0 or one of the two semi-lines
z = 0 or z = L at t 0. We indicate this intersection point by (t¯, z̄). The corresponding
value of A, call it A, is positive by hypothesis. From the continuity equation, A satisfies
along the line l the following ordinary differential equation:
dA ∂ ū
= −A ,
dt ∂z
where here the dA/dt indicates the directional derivative along l. Therefore,
t∗
∗ ∗ ∂ ū
A(t , z ) = A τ, zu (τ ) dτ > 0,
t¯ ∂z
in contradiction with the hypothesis. Therefore, we must have A > 0.
S ECTION 20 101
Here we derive now an a priori estimate for the solution of system (20.18) under
the hypotheses of α = 1, sub-critical smooth flow, and A > 0. We will consider the
following initial and boundary conditions:
initial conditions A(0, z) = A0 (z), Q(0, z) = Q0 (z), z ∈ (0, L) (20.37)

boundary conditions W1 (t, 0) = g1 (t), t ∈ I, (20.38)
W2 (t, L) = g2 (t), t ∈ I.
Let the quantity e be defined as
ρ 2
e= Aū + Ψ, (20.39)
2
where Ψ = Ψ (A) is given by
A
Ψ (A) = ψ(ζ ) dζ. (20.40)
A0
Here and in what follows we omit to indicate the dependence of ψ on A0 and β, since
it is not relevant to obtain the desired result, which can be however extended also to the
general case where the coefficients A0 and β depend on z.
An energy of the 1D model is given by
L
E(t) = e(t, z) dz, t ∈ I. (20.41)
0
Indeed, owing to the assumptions we have made on ψ in (20.14), we may observe
that ψ attains a minimum at A = A0 , since
Ψ (A0 ) = Ψ (A0 ) = 0 and Ψ (A) > 0, ∀A > 0.

It follows that Ψ (A) 0, ∀A > 0. Consequently, E(t) is a positive function for all Q
and A > 0 and, moreover,

E(t) = 0 iff A(t, z), Q(t, z) = (A0 , 0), ∀z ∈ (0, L).
The following lemma holds.
L EMMA 20.4. In the special case α = 1, system (20.12), supplied with an algebraic
pressure-area relationship of the form (20.13) and under conditions (20.14), satisfies
the following conservation property, ∀t ∈ I :
t L t L

E(t) + ρKR ū dz dτ +
2
Q(Ptot − Pext ) dτ = E(0), (20.42)
t0 0 t0 0
where E(0) depends only on the initial data A0 and Q0 , while Ptot = P + 12 ρ ū2 is the
fluid total pressure.
P ROOF. Let us multiply the second equation of (20.12) by ū and integrate over (0, L).
We will analyse separately the four terms that are obtained.
• First term:
L L L
∂(Aū) 1 ∂ ū2 ∂A 2
I1 = ū dz = A dz + ū dz
0 ∂t 2 0 ∂t 0 ∂t
L L
1 d 1 ∂A
= Aū2 dz + ū2 dz. (20.43)
2 dt 0 2 0 ∂t
• Second term:
L L L
∂(Aū2 ) ∂(Aū) 2 ∂ ū
I2 = α ū dz = α ū dz + Aū2 dz
0 ∂z 0 ∂z 0 ∂z
L
1 ∂(Aū) 2 1 L ∂A 3 3 L 2 ∂ ū
=α ū dz + ū dz + Aū dz
2 0 ∂z 2 0 ∂z 2 0 ∂z
L
1 ∂Q 2 1 L ∂(Aū3 )
=α ū dz + dz . (20.44)
2 0 ∂z 2 0 ∂z
Now, using the continuity equation, we obtain
L
α ∂A 2 3 L
I2 = −
ū dz + Aū 0 . (20.45)
2 0 ∂t
• Third term:
L
A ∂P 1 L ∂
I3 = ū dz = A (P − Pext )ū dz
0 ρ ∂z ρ 0 ∂z
L
1 ∂Q L
= − ψ(A) dz + (P − Pext )Q0 . (20.46)
ρ 0 ∂z
Again, using the first of (20.12), we have

L
1 ∂A L
I3 =
ψ(A) dz + (P − Pext )Q 0
ρ 0 ∂t

1 d L L
= Ψ (A) dz + (P − Pext )Q0 .
ρ dt 0
• Fourth term:
L L
Q
I4 = Kr ū dz = Kr ū2 dz. (20.47)
0 A 0
By summing the four terms and multiplying by ρ, we obtain the following equality
when α = 1:
L
1 d L 2 d L L
ρ Aū dz + Ψ (A) dz + ρKr ū2 dz + Q(Ptot − Pext )0 = 0.
2 dt 0 dt 0 0
(20.48)
Integrating Eq. (20.48) in time between t0 and t leads to the desired result.
S ECTION 20 103
In order to draw an energy inequality from (20.42), we need to investigate the sign of
the last term on the left-hand side. With this aim, let us first analyse the homogeneous
case g1 = g2 = 0.
We will rewrite the boundary term in (20.42) as a function of A, ψ(A) and c1 (which,
in its turn, depends on A, see (20.17)).
If g1 = g2 = 0 in (20.38), then
A A
c1 (ζ ) c1 (ζ )
at z = 0, W1 = ū + dζ = 0 ⇒ ū(t, 0) = − dζ,
A0 ζ A0 ζ
A A
c1 (ζ ) c1 (ζ )
at z = L, W2 = ū − dζ = 0 ⇒ ū(t, L) = dζ
A0 ζ A0 ζ
and thus
L
Q(Ptot − Pext )0 = F A(t, 0) + F A(t, L) , (20.49)
where
A A 2
c1 (ζ ) 1 c1 (ζ )
F (A) = A dζ ψ(A) + ρ dζ . (20.50)
0 ζ 2 A0 ζ
From our assumption of sub-critical flow we have |ū| < c1 which implies that at z = 0
and z = L we have
A
c1 (ζ )
dζ < c1 (A). (20.51)
ζ
A0
We are now in the position to conclude with the following result.
L EMMA 20.5. If the function pressure-area relationship P = ψ(A) is such that

F (A) > 0 whenever (20.51) is satisfied, then inequality
t L
E(t) + ρKr ū2 dz dτ E(0) (20.52)
t0 0
holds for system (20.12), provided homogeneous conditions on the characteristic vari-
ables, W1 = 0 and W2 = 0, are imposed at z = 0 and z = L, respectively.
P ROOF. It is an immediate consequence of (20.42), (20.49) and (20.50).
By straightforward computations, one may verify that the pressure–area relationship

given in (20.16) satisfies the hypotheses of Lemma 20.5 (see F ORMAGGIA , G ERBEAU ,
N OBILE and Q UARTERONI [2001]). Therefore, in that case the 1D model satisfies the
energy inequality (20.52).
Under relation (20.16), we can prove a more general energy estimate, valid also in
the case of non homogeneous boundary conditions. We state the following result.
L EMMA 20.6. If the pressure–area relationship is given by (20.16), and the boundary
data satisfy
g1 (t) > −4c1,0(0) and g2 (t) < 4c1,0 (L), ∀t ∈ I, (20.53)
where

β0 (z)
c1,0 (z) = A0 (z)−1/4
ρ
is the value of c1 at the reference vessel area, then there exists a positive quantity G(t)
which continuously depends on the boundary data g1 (t) and g2 (t), as well as on the
values of the coefficients A0 and β, at z = 0 and z = L, such that, for all t ∈ I ,
t L t
E(t) + ρKr ū dz dτ E(0) +
2
G(t) dt. (20.54)
t0 0 0
P ROOF. We will consider only the case where g1 = 0 and g2 = 0, since the most general
case may be derived in a similar fashion. We recall that relationship (20.16) together
with the assumption of sub-critical flow, complies with the conditions stated for F (A)
in Lemma 20.5. Then from (20.48) we obtain the following inequality:
L
d
E + ρKr ū dz Q(Ptot − Pext )
2
dt 0 z=0

1
A|ū|ψ(A) + ρA|ū|3 . (20.55)
2 z=0
At z = 0, we have from (20.33) that
ū + 4(c1 − c1,0 ) = g1 .
On the other hand, the condition λ1 = ū + c1 > 0 gives
1
c1 < (g1 + 4c1,0). (20.56)
3
Since c1 is a non-negative quantity, we must necessarily have g1 > −4c1,0 . We now
note that from (20.16) and the definition of c1,0 we may write

ψ(A) = 2ρ c12 (A) − c1,0
2
,
which together with (20.56) and the fact that c1,0 is a positive function, allows us to
state that, at z = 0,
2ρ 2
ψ(A) g1 + 15c1,0
2
+ 8g1 c1,0 ≡ f1 (g1 ), (20.57)
9
where f1 is a positive continuous function depending parametrically on the values of
A0 and β0 at z = 0. Furthermore, condition |ū| < c1 together with inequality (20.56)
imply that
|ū| f2 (g1 ),
S ECTION 20 105
being f2 another positive and continuous function. Finally, from the definition of ψ and
c1,0 we have
A20 2 2A20 2
A= ψ(A) + A0 ψ (A) + A0
β02 2
β0
2A20 2
f1 (g1) + A 0 ≡ f3 (g1 ),
β02
where we have exploited (20.57). By combining all previous inequalities, we deduce
that the right-hand side in (20.55) may be bounded by a positive and continuous function
of the boundary data g1 that depend parametrically on the value of A0 and β0 at z = 0.
By repeating a similar argument for the boundary conditions at z = L, we then obtain
the desired stability inequality.
20.3.4. Weak form

We consider the hyperbolic system (20.20) with initial condition U = U0 , at t = t0 , and
appropriate boundary conditions at z = 0 and z = L. We indicate by C01 ((0, L)×[t0 , t1 ))
the set of functions which are the restriction to (0, L) × [t0 , t1 ) of C 1 functions with
compact support in (0, L) × (−∞, t1 ). We will assume that U0 is a bounded measurable
function in (0, L).
A function U ∈ [L∞ ((0, L) × [t0 , t1 ))]2 is a weak solution of the equation in conser-
vation form (20.20) if for all φ ∈ [C01 ((0, L) × [t0 , t1 ))]2 we have
t 1 L L
∂φ ∂φ
U· + F (U ) · − S(U ) · φ dz dt + U0 · φ|t =0 = 0. (20.58)
t0 0 ∂t ∂z 0
Moreover, we will require that U complies given boundary conditions.
A solution of (20.58) is called a weak solution of our hyperbolic system. Clearly,
“classical” smooth solutions of (20.20) are also weak solutions. Conversely, it may be
shown that a smooth weak solution, i.e., belonging to [C 1 ((0, L) × [t0 , t1 ))]2 , is also
solution of (20.20) in a classical sense. However, the weak form accommodates also
for less regular U . In particular, weak solutions of our hyperbolic problem may be
discontinuous. The weak form is furthermore the basis of a class of numerical schemes,
in particular, the finite element method, as already seen for the Navier–Stokes equations.
R EMARK 20.7. The conservation formulation (20.20) accounts also for mechanical
properties which vary smoothly along z. However, there are some fundamental dif-
ficulties in extending it to the case of discontinuous mechanical characteristics (e.g.,
discontinuous β). On the other hand, this situation has a certain practical relevance, for
instance in stented arteries or in the presence of a vascular prosthesis. A stent is a metal
meshed wire structure inserted into a stenotic artery (typically a coronary) by angio-
plasty, in order to restore the original lumen dimension. Vascular prostheses are used to
treat degenerative pathologies, such as aneurysms, or when angioplasty is not possible.
A possibility (F ORMAGGIA , N OBILE and Q UARTERONI [2002]) is to model the
sharp variation of the Young modulus at the interface between the artery and the prosthe-
sis by a regular function. Fig 20.2 illustrates a possible description of the change in the
F IG . 20.2. The sharp variation of the Young modulus E from the value E0 to the value E1 , due to the
presence of a prosthesis, is modeled by a smooth function. One may argue what would happen when the
parameter δ in figure tends to zero.
Young modulus due to the presence of a prosthesis. One may argue what would happen
when the parameter δ in figure tends to zero. Numerical experiments have shown that
the solution remains bounded although it becomes discontinuous at the location of the
discontinuity in the Young modulus. This fact has been recently investigated in C ANIC
[2002] where an expression for the jump of mass flow and area across the discontinuity
is derived by computing a particular limit of weak solutions of a regularised problem.
More details are found in the cited reference.
20.3.5. An entropy function for the 1D model

Let us consider the hyperbolic system written in quasi-linear form (20.18). A pair of
functions e : R2 → R and Fe : R2 → R is called entropy pair for the system if e is a
convex function of U (called entropy) and the following condition is satisfied:

de T ∂Fe
H(U ) = (20.59)
dU ∂U
for all admissible values of U .
Fe is the entropy flux associated to the entropy e. If the hyperbolic system admits an
entropy pair then the entropy function satisfies a conservation law of the form
∂e ∂Fe
+ + Be (U ) = 0,
∂t ∂z
where
de ∂Fe dA0 ∂Fe dβ
Be (U ) = · B(U ) − −
dU ∂A0 dz ∂β dz
is a source term. The last two terms in the previous expression account for the possible
dependence of the coefficients A0 and β on z.
The existence of an entropy pair is of a certain importance when studying the weak
solution of the hyperbolic problem and, in particular, discontinuous solutions (more
S ECTION 20 107
details in L AX [1973] and G ODLEWSKI and R AVIART [1996]). Although we have here
considered only smooth solutions, the identification of an entropy for our problem is
important to set the basis for the extension of the model to more general situations.
In the case α = 1,
1 1 Q2
e = ρAū2 + Ψ (A) = ρ + Ψ (A)
2 2 A
is indeed an entropy for the problem at hand, with associated flux

1
Fe = Q ψ(A) + ρ ū2 = Q(Ptot − Pext ).
2
Indeed, we have
ρ ū2 ∂ψ
Q ∂A (A) − ρ ū3
= − 2 + ψ(A) ,
∂e ∂Fe
=
∂U ρ ū ∂U ψ(A) + 32 ρ ū2
and we may directly verify condition (20.59) by recalling (20.19). Furthermore, Be =
ρKr ū2 and the entropy balance equation thus read

∂ 1 ∂ 1
ρAū2 + Ψ (A) + Q ψ(A) + ρ ū2 + ρKr ū2 = 0. (20.60)
∂t 2 ∂z 2
It is valid for any smooth solution of our hyperbolic model. Furthermore, the following
lemma ensures the convexity of e.
L EMMA 20.7. The entropy

ρ Q2
e(A, Q) = + Ψ (A)
2 A
is convex for all A > 0.
P ROOF. By a straightforward calculation one finds that the Hessian of e is given by

 2 
∂ e ∂ 2e

 ∂A2 ∂A∂Q  ρ ū2 + c12 −ū

He =  2  = .
∂ e ∂ 2e  A −ū 1
∂A∂Q ∂Q2
Its eigenvalues are
%
c12 + ū2 + 1 ± (c12 + ū2 + 1)2 − 4c12
λ1,2 (He ) = ρ .
2A
The condition for the discriminant to be positive is
2
4c12 c12 + ū2 + 1 .
Since c1 > 0 whenever A > 0, this inequality is equivalent to impose that
c12 + ū2 + 1 − 2c1 = (c1 − 1)2 + ū2 0,
which it is always true. Therefore, the two eigenvalues are strictly positive for all A > 0.
This completes our proof.
20.4. More complex wall laws that account for inertia and viscoelasticity
The algebraic relation (20.13) assumes that the wall is instantaneously in equilibrium
with the pressure forces acting on it. Indeed, this approach correspond to the indepen-
dent ring model introduced in Section 15.
At the price of some approximations it is possible to maintain the simple structure of
a two-equations system while introducing effects, such as the inertia, which depend on
the time derivative of the wall displacement.
We will consider as starting point relation (16.17) where we account for the inertia
term and we model the viscoelastic property of the wall by adding a term proportional to
the displacement rate, while we will still use the approximation (16.15) for the forcing
term. We may thus write
∂ 2η ∂η
P − Pext = γ0 + γ1 + ψ(A; A0 , β), (20.61)
∂t 2 ∂t
where γ0 = ρw h0 , γ1 = γ /R02 and the last term is the elastic response, modelled is the
same way as done before. Here γ is the same viscoelasticity coefficient of (16.17) and
η is the wall displacement, linked to A by (20.2).
In the following, we indicate by Ȧ and Ä the first and second time derivative of A.
We will substitute the following identities:

∂η 1 ∂ 2η −1/2 1 1
= √ Ȧ, = π √ Ä − √ Ȧ 2
,
∂t 2 πA ∂t 2 2 A 4 A3
that are derived from (20.2), into (20.61) to obtain a relation that links the pressure also
to the time derivatives of A, which we write in all generality as
P − Pext = ψ̃(A, Ȧ, Ä; A0 ) + ψ(A; A0 , β),

where ψ̃ is a non-linear function which derives from the treatment of the terms con-
taining the time derivative of η. Since it may be assumed that the contribution to the
pressure is in fact dominated by the term ψ, we will simplify this relationship by lin-
earising ψ̃ around the state A = A0 , Ȧ = Ä = 0. By doing that, after some simple
algebraic manipulations, one finds
γ0 γ1
P − Pext = √ Ä + √ Ȧ + ψ(A; A0 , β). (20.62)
2 πA0 2 πA0
Replacing this expression for the pressure in the momentum equation requires to
compute the term
A ∂P γ0 A ∂ 3A γ1 A ∂ 2 A A ∂ψ
= √ 2
+ √ + .
ρ ∂z 2ρ πA0 ∂z∂t 2ρ πA0 ∂z∂t ρ ∂z
The last term in this equality may be treated as previously, while the first two terms may
be further elaborated by exploiting the continuity equation. Indeed, we have
∂ 2A ∂ 2Q ∂ 3A ∂ 3Q
=− 2 , = − .
∂z∂t ∂z ∂z∂t 2 ∂t∂z2
S ECTION 20 109
Therefore, the momentum equation with the additional terms deriving from inertia
and viscoelastic forces becomes
∂Q ∂F2 γ0 A ∂ 3Q γ1 A ∂ 2 Q
+ − √ − √ + S2 = 0, (20.63)
∂t ∂z 2ρ πA0 ∂t∂z2 2ρ πA0 ∂z2
where with F2 and S2 we have indicated the second component of F and S, respectively.
R EMARK 20.8. This analysis puts into evidence that the wall inertia introduces a dis-
persive term into the momentum equation, while the viscoelasticity has a diffusion ef-
fect.
20.5. Some further extensions
More general one-dimensional models may be derived by accounting for vessel curva-
ture. This may be accomplished by enriching the description of the velocity field on
each vessel section to allow asymmetries of the velocity profile to develop.
Another enhancement of the model is to account for vessel branching. By employ-
ing domain decomposition techniques, each branch is simulated by a separate one-
dimensional model and interface conditions are used to account for the appropriate
“transfer” of mass and momentum across the branching point. All these aspects are
not covered in these notes. They are subject of current research and preliminary results
may be found in F ORMAGGIA , L AMPONI and Q UARTERONI [2003].
Beside providing valuable information about average pressure and mass flux along
an arterial segment, a one-dimensional model of blood flow may be used in the con-
text of a multiscale/multimodel description of the cardiovascular system. In the multi-
scale framework, models of different level of complexity of the various cardiovascular
elements are coupled together with the objective of simulating the whole cardiovascu-
lar system. Only the elements of major interest for the problem under study will be
simulated at the highest level of detail (e.g., by employing a three-dimensional fluid–
structure interaction model), while reduced models are adopted in the remaining parts.
This technique allows us to account (at least partially) for the complex feedback mech-
anisms of the complete cardiovascular system, while keeping the overall computational
costs at a reasonable level. More details on this technique may be found in F ORMAG -
GIA , N OBILE , Q UARTERONI and V ENEZIANI [1999], F ORMAGGIA , G ERBEAU , N O -
BILE and Q UARTERONI [2001], Q UARTERONI , R AGNI and V ENEZIANI [2001] while
in P IETRABISSA , Q UARTERONI , D UBINI , V ENEZIANI , M IGLIAVACCA and R AGNI
[2000] a first example on the use of this multiscale approach for a realistic clinical ap-
plication is presented.
C HAPTER VII
21. Some numerical results
We provide some numerical results to illustrate applications of the techniques dis-

cussed in the previous sections. The aim here is to show the potential of the numerical
modelling to reproduce realistic flow fields relevant for medical investigations. Many
of the results here presented are substantially taken from previous works of the au-
thors, in particular from Q UARTERONI , T UVERI and V ENEZIANI [2000], F ORMAG -
GIA , G ERBEAU , N OBILE and Q UARTERONI [2001] and F ORMAGGIA , N OBILE and
Q UARTERONI [2002]. More details and other examples may be found in the cited ref-
erences.
21.1. Compliant pipe
Here we consider two examples of a fluid–structure interaction problem like the one
presented in Section 19, namely a 2D and a 3D computation of a pressure wave in a
compliant tube.
In the 2D case, we have considered a rectangular domain of height 1 cm and
length L = 6 cm. The fluid is initially at rest and an over pressure of 15 mmHg
(2 · 104 dynes/cm2) has been imposed at the inlet for 0.005 seconds. The viscosity
of the fluid is equal to 0.035 poise, its density is 1 g/cm3, the Young modulus of the
structure is equal to 0.75 · 106 dynes/cm2, its Poisson coefficient is 0.5, its density is
1.1 g/cm3 and its thickness is 0.1 cm.
In the 3D case, our computation has been made on a cylindrical domain of ra-
dius R0 = 0.5 cm and length L = 5 cm, with the following physical parameters:
fluid viscosity: 0.03 poise, fluid density: 1 g/cm3 , Young modulus of the structure:
3 · 106 dynes/cm2, Poisson coefficient: 0.3 and structure density: 1.2 g/cm3 . Again, an
over-pressure of 10 mmHg (1.3332 · 104 dynes/cm2) is imposed at the inlet for 0.005
seconds.
The fluid equations are solved using the ALE approach, with a piece-wise linear finite
element space discretisation. More precisely, for the 2D case the pressure is piece-wise
linear on triangular elements and the velocity is linear over each of the four sub-triangles
obtained by joining the midpoints of the edges of each pressure triangle (this is the so
called P1isoP2–P1 discretisation). We have employed the Yosida technique illustrated in
Section 14.5.2. For the 3D case we have used a stabilised scheme (H UGHES , F RANCA
and BALESTRA [1986]) and piece-wise linear elements for both velocity and pressure.
For the 2D case, the equation for the structure displacement (18.13) has been solved
using a P 1 finite element space discretisation, with nodes coincident with the ones of
111
112 A. Quarteroni and L. Formaggia C HAPTER VII
the pressure discretisation. In the 3D case, we have used a shell-type formulation (S IMO
and F OX [1989], S IMO , F OX and R IFAI [1989]) to describe the dynamics of the wall
structure. In both cases, the coupling scheme adopts a sub-iterations strategy of the type
illustrated in Section 19.
In order to reduce spurious wave reflections at the outlet, we have coupled the fluid–
structure interaction problem with a one-dimensional system of the type described in
Section 20. For more details on this technique see F ORMAGGIA , G ERBEAU , N OBILE
and Q UARTERONI [2001], as well as F ORMAGGIA , G ERBEAU , N OBILE and Q UAR -
TERONI [2002].
Figs. 21.1 and 21.2 show the fluid pressure and the domain deformation in the 2D and
the 3D case, respectively. For the sake of clarity, the displacements shown in Fig. 21.2
are magnified by a factor 10.
F IG . 21.1. Pressure pulse entering at the inflow. A non-reflecting boundary condition at the outlet has been
obtained by the coupling with a 1D hyperbolic model. Solutions every 5 ms.
S ECTION 21 113
F IG . 21.2. A pressure pulse traveling in a 3D compliant vessel. The displacement of the structure has been
magnified by a factor 10. A non-reflecting boundary condition at the outlet has been obtained by the coupling
with a 1D hyperbolic model (not shown in the picture).
21.2. Anastomosis models
Anastomosis is the a surgical operation by which the functionality of a blocked artery

(typically a coronary) is restored thanks to by-pass. The flow condition when the blood
in the by-pass re-joins the main artery may be critical. If we have a large recirculation
F IG . 21.3. A model of a coronary by-pass anastomosis (top) and the velocity vector field on the median plane
at four different instants of the heart beat. Flow at systole (top, left), initial deceleration phase (top, right),
beginning of diastole (bottom, left) and end of diastole (bottom, right). The recirculation regions upstream
and downstream of the junction are evident.
area, the higher latency time of blood particles there may favor plaque growing and
cause a new blockage further downstream.
S ECTION 21 115
The simulations here presented aim at highlight the problem. We illustrate the flow
in the median plane of a 3D model of an anastomosis.7 The junction angle is 15 de-
grees. The diameter of the occluded branch (below) is 1 cm, and the one of the by-pass
(above) is 0.96 cm. The simulations have been carried out setting the dynamic viscos-
ity µ = 0.04 g cm−1 s−1 and the density ρ = 1 g cm−3 . In this simulation the vessel
wall has been assumed fixed and the boundary conditions prescribe null velocity on the
walls and on the upstream section of the stenotic branch (100% stenosis), while a par-
abolic velocity profile has been prescribed at the inlet section with a peak velocity of
56 cm s−1 , corresponding to a flow rate of 1320 ml min−1 . On the downstream section
a Neumann-type condition has been assigned.
Fig. 21.3 clearly illustrates the appearance and the evolution of the flow recirculation
zones during the different phases of the heart beat.
21.3. Pressure wave modification caused by a prosthesis
Here we present a numerical simulation obtained using the one-dimensional model

(20.12) to investigate the effect of a prosthesis in an artery, in particular with respect
to the alteration of the pressure wave pattern. To that purpose we have considered the
portion of an artery of length L and a prosthesis of length l (see Fig. 21.4) and a Young
modulus varying as already illustrated in Fig. 20.2.
In order to assess the effect of the changes in vessel wall elastic characteristic on the
pressure pattern, we have devised several numerical experiments. Two types of pressure
input have been imposed at z = 0, namely an impulse input, that is a single sine wave
with a small time period and a single sine wave with a more realistic time period (see
Fig. 21.5). The impulse has been used to better highlight the reflections induced by the
vascular prosthesis.
The part that simulates the presence of the prosthesis or stent of length L is comprised
between coordinates a1 and a2 . The corresponding Young’s modulus has been taken as
a multiple of the basis Young’s modulus E0 associated to the physiological tissue.
F IG . 21.4. The layout of our numerical experiment. The points P , M and D are used as ‘monitoring stations’
to assess the modifications on the pressure wave caused by the prosthesis.
7 The model geometry has been provided by the Vascular Surgery Skejby Sygheus of the Aahrus University
Hospital in Denmark.
F IG . 21.5. The two types of pressure input profiles used in the numerical experiments: an impulse (left) and
a more realistic sine wave (right).
TABLE 21.1
Data used in the numerical experiments
Parameters Value
Fluid Input pressure amplitude 20 × 103 dyne/cm2

Viscosity, ν 0.035 poise
Density, ρ 1 g/cm3
Structure Young’s modulus, E0 3 × 106 dyne/cm2
Wall thickness, h 0.05 cm
Reference radius, R0 0.5 cm
Three locations along the vessel have been identified and indicated by the letters
D (distal), M (medium) and P (proximal). They will be taken as monitoring point
for the pressure variation. Different prosthesis length L have been considered; in all
cases points P and D are located outside the region occupied by the prosthesis. Ta-
ble 21.1 indicates the basic data which have been used in all numerical experiments.
In this numerical experiment we have considered the conservation form (20.20) set-
ting the friction term Kr to zero. The numerical scheme adopted is a second order
Taylor–Galerkin (D ONEA , G IULIANI , L AVAL and Q UARTAPELLE [1984]). A time step
t = 2 × 10−6 s and the initial values A = A0 and Q = 0 have been used throughout.
At the outlet boundary z = L we have kept W2 constant and equal to its initial value
(non-reflecting boundary condition). At the inlet boundary we have imposed the chosen
pressure input in an approximate fashion, following a technique of the type illustrated
in Section 20.3.2.
21.3.1. Case of an impulsive pressure wave

In Fig. 21.6 we show the results obtained for the case of a pressure impulse. We compare
the results obtained with uniform Young modulus E0 and the corresponding solution
when E1 = 100E0 , l = 5 cm and the transition zone between healthy artery and pros-
S ECTION 21 117
F IG . 21.6. Pressure history at points P , M and D of Fig. 21.4, for an impulsive input pressure, in the case
of constant (upper) and variable (lower) E.
thesis is δ = 0.5 cm. We have taken L = 15 cm and a non-uniform mesh of 105 finite
elements, refined around the points a1 and a2 . When the Young modulus is uniform,
the impulse travels along the tube undisturbed. The numerical solution shows a little
dissipation and dispersion due to the numerical scheme. In the case of variable E the
situation changes dramatically. Indeed, as soon as the wave enters the region at higher
Young’s modulus it gets partially reflected (the reflection is registered by the positive
pressure value at point P and t ≈ 0.015 s) and it accelerates. Another reflection oc-
curs at the exit of the ‘prosthesis’, when E returns to its reference value E0 . The point
M indeed registers an oscillatory pressure which corresponds to the waves that are re-
flected back and forth between the two ends of the prosthesis. The wave at point D is
much weaker, because part of the energy has been reflected back and part of it has been
‘captured’ inside the prosthesis itself.
21.3.2. Case of a sine wave

Now, we present the case of the pressure input given by the sine wave with a larger
period shown in Fig. 21.5, which describes a situation closer to reality than the im-
pulse. We present again the results for both cases of a constant and a variable E. All
other problem data have been left unchanged from the previous simulation. Now, the
interaction among the reflected waves is more complex and eventually results in a less
oscillatory solution (see Fig. 21.7). The major effect of the presence of the stent is a
pressure increase at the proximal point P , where the maximum pressure is approxi-
mately 2500 dynes/cm2 higher than in the constant case. At a closer inspection one
may note that the interaction between the incoming and reflected waves shows up in
discontinuities in the slope, particularly for the pressure history at point P . In addition,
the wave is clearly accelerated inside the region where E is larger.
In Table 21.2 we show the effect of a change in the length of the prosthesis by com-
paring the maximum pressure value recorded for a prosthesis of 4, 14 and 24 cm, respec-
tively. The values shown are the maximal values in the whole vessel, over one period.
Here, we have taken L = 60 cm, δ = 1 cm, a mesh of 240 elements and we have posi-
tioned in the three cases the prosthesis in the middle of the model. The maximum value
is always reached at a point upstream the prosthesis. In the table we give the normalised
distance between the upstream prosthesis section and of the point where the pressure
attains its maximum.
Finally, we have investigated the variation of the pressure pattern due to an increase
of k = E/E0 . Fig. 21.8 shows the result corresponding to L = 20 cm and δ = 1 cm
and various values for k. The numerical result confirms the fact that a stiffer prosthesis
TABLE 21.2
Maximum pressure value for prosthesis of different length
Prosthesis length Maximal pressure Maximum location

(cm) (dyne/cm2 ) zmax / l
4 23.5 × 103 0.16

14 27.8 × 103 0.11
24 30.0 × 103 0.09
S ECTION 21 119
F IG . 21.7. Pressure history at points P , M and D of Fig. 21.4, for a sine wave input pressure, in the case of
constant (upper) and variable (lower) E.
F IG . 21.8. Pressure history at point P of Fig. 21.4, for a sine wave input pressure and different Young’s
moduli E = kE0 .
causes a higher excess pressure in the proximal region, a fact that may have negative
effects on the heart.
21.4. Some examples of the geometrical multiscale approach
We end this section by giving some examples of the geometrical multiscale approach,
where models of different geometrical complexity are coupled together to provide the
simulation of the global cardiovascular system, at different level of detail.
Fig. 21.9 shows an example of the simulation of a by-pass, with the interplay be-
tween three-dimensional, one-dimensional and lumped parameters models. A detailed
description of the flow in the by-pass is obtained by solving the fluid–structure interac-
tion problem (here using a two-dimensional model). The presence of the global cardio-
vascular system is provided by a system of algebraic and ordinary differential equations
(ODE) for average mass flow and pressure. This system is here illustrated by means of
an electrical analog, where voltage plays the role of average pressure and the current
that of mass flow. A transition between the two models is provided by the use of the
one-dimensional description detailed in the previous section.
A simpler example of this coupling strategy, yet on a realistic three-dimensional
geometry, is shown in Fig. 21.10. A three-dimensional model of the modified Blalock–
Taussig shunt a surgical operation meant to cure the consequences of a severe car-
diac malformation, has been devised with the intent of finding the optimal design for
the shunt. The three-dimensional model (on a fixed geometry) has been coupled with
S ECTION 21 121
F IG . 21.9. On top we show a global model of the circulatory system where a coronary by-pass is being
simulated by a Navier–Stokes fluid–structure interaction model. The rest of the circulatory system is described
by means of a lumped parameter model, based on the solution of a system of ODEs, is here represented by an
electrical circuit analog in the bottom part of the figure.
the systemic lumped parameter model, which provides the boundary conditions for the
Navier–Stokes equations at the inlet and outlet sections. Thanks to this multiscale ap-
proach it has been possible to compute velocity profiles and flow patterns which are
F IG . 21.10. Simulation of the haemodynamics in the modified Blalock–Taussig shunt obtained using a geo-
metrical multiscale approach. Velocity field in the ascending aorta at two different times during the cardiac
cycle.
closer to reality than those obtained by using more standard boundary conditions. An
analysis of the technique is found in Q UARTERONI and V ENEZIANI [2003], while more
details on this and other test cases may be found in P IETRABISSA , Q UARTERONI ,
D UBINI , V ENEZIANI , M IGLIAVACCA and R AGNI [2000], M IGLIAVACCA , L AGANÁ ,
P ENNATI , DE L EVAL , B OVE and D UBINI [2004].
22. Conclusions
The development of mathematical models, algorithms and numerical simulation tools

for the investigation of the human cardiovascular system has received a great impulse in
the last years. These notes intended to cover just a few of the relevant issues. There are
however other important aspects which require the use of sophisticated mathematical
and numerical tools. We here mention just a few, namely the reconstruction of geome-
tries from medical data; the transport of biochemicals in blood and vessel wall tissue;
the heart dynamics; blood rheology. Besides, the need of validating the models calls for
development of accurate in-vivo measurement techniques.
The number and complexity of the mathematical, numerical and technological prob-
lems involved makes the development of tools for accurate, reliable and efficient simu-
lations of the human cardiovascular system one of the challenges of the next decades.
Acknowledgements
The authors thank Prof. Alessandro Veneziani and Dr. Fabio Nobile for their valuable
contributions during the preparation of these notes and for having provided most of the
numerical results here presented. We thank Dr. G. Dubini and Dr. F. Migliavacca for the
availability of the numerical results for the modified Blalock–Taussig shunt.
Our research activity on the mathematical modelling of the cardiovascular system has
been partially supported by grants from various research agencies, which we gratefully
acknowledge. In particular, grants 21-54139.98, 21-59230.99 and 20-61862.00 from
the Swiss National Science Foundation, the project of Politecnico di Milano “LSC-
S ECTION 22 123
Multiscale Computing in Biofluiddynamics”, the project “Agenzia-2000” by the Ital-

ian CNR, titled “Modeling the fluid structure interaction in the arterial system”, and a
research contract “Cofin-2000” by the Italian Ministry of Education (MURST) titled
“Scientific Computing: Innovative Models and Numerical Methods”. Finally, the au-
thors acknowledge the support by the European Union through the Research Training
and Network project “HaeMOdel”, contract number HPRN-CT-2002-002670.
References
A CHDOU , Y., G UERMOND , J.L. (2000). Convergence analysis of a finite element projection/Lagrange–
Galerkin method for the incompressible Navier–Stokes equations. SIAM J. Numer. Anal. 37 (3), 799–826.
A RIS , R. (1962). Vectors, Tensors and the Basic Equations of Fluid Mechanics (Prentice Hall, New York).
BARNARD , A.C.L., H UNT , W.A., T IMLAKE , W.P., VARLEY , E. (1966). A theory of fluid flow in compliant
tubes. Biophys. J. 6, 717–724.
B EIRÃO DA V EIGA , H. (2004). On the existence of strong solutions to a coupled fluid–structure evolution
problem. J. Math. Fluid Mech. 6, 21–52.
B OUKIR , K., M ADAY, Y., M ÉTIVET, B., R AZAFINDRAKOTO , E. (1997). A high-order characteristics/finite
element method for the incompressible Navier–Stokes equations. Internat. J. Numer. Methods Flu-
ids 25 (12), 1421–1454.
B REZIS , H. (1983). Analyse Fonctionnelle (Masson, Paris).
B REZZI , F., F ORTIN , M. (1991). Mixed and Hybrid Finite Elements, Springer Ser. Comput. Math. 5
(Springer-Verlag, Berlin).
C ANIC , S. (2002). Blood flow through compliant vessels after endovascular repair: wall deformations induced
by the discontinuous wall properties. Comput. Visual. Sci. 4 (3), 147–155.
C ANIC , S., K IM , E. (2003). Mathematical analysis of the quasilinear effects in a hyperbolic model of blood
flow through compliant axi-symmetric vessels. Math. Methods Appl. Sci. 26 (14), 1161–1186.
C HORIN , A.J., M ARSDEN , J.E. (1990). A Mathematical Introduction to Fluid Mechanics, third ed., Texts
Appl. Math. 4 (Springer-Verlag, New York).
C IARLET , P.G. (1988). Mathematical Elasticity, Volume I: Three-Dimensional Elasticity, Stud. Math. Appl.
20 (North-Holland, Amsterdam).
C IARLET , P.G. (1998). Introduction to Linear Shell Theory (Gauthier–Villars, Paris).
C IARLET , P.G. (2000). Mathematical Elasticity, Volume III: Theory of Shells (North-Holland, Amsterdam).
C OKELET , G.R. (1987). The rheology and tube flow of blood. In: Skalak, R., Chen, S. (eds.), Handbook of
Bioengineering (McGraw–Hill, New York).
D ONEA , J., G IULIANI , S., L AVAL , H., Q UARTAPELLE , L. (1984). Time-accurate solutions of advection–
diffusion problems by finite elements. Comput. Methods Appl. Mech. Engrg. 45, 123–145.
D UVAUT , G., L IONS , J.-L. (1976). Inequalities in Mechanics and Physics (Springer-Verlag, Berlin).
FARHAT , C., L ESOINNE , M. (2000). Two efficient staggered algorithms for the serial and parallel solution
of three-dimensional nonlinear transient aeroelastic problems. Comput. Methods Appl. Mech. Engrg. 182,
499–515.
FARHAT , C., L ESOINNE , M., M AMAN , N. (1995). Mixed explicit/implicit time integration of coupled
aeroelastic problems: three-field formulation, geometry conservation and distributed solution. Internat.
J. Numer. Methods Fluids 21, 807–835.
F ORMAGGIA , L., G ERBEAU , J.-F., N OBILE , F., Q UARTERONI , A. (2001). On the coupling of 3D and
1D Navier–Stokes equations for flow problems in compliant vessels. Comput. Methods Appl. Mech. En-
grg. 191, 561–582.
F ORMAGGIA , L., G ERBEAU , J.-F., N OBILE , F., Q UARTERONI , A. (2002). Numerical treatment of defective
boundary conditions for Navier–Stokes equations. SIAM J. Numer. Anal. 40 (1), 376–401.
F ORMAGGIA , L., L AMPONI , D., Q UARTERONI , A. (2003). One-dimensional models for blood flow in arter-
ies. J. Engrg. Math. 47, 251–276.
F ORMAGGIA , L., N OBILE , F. (1999). A stability analysis for the Arbitrary Lagrangian Eulerian formulation
with finite elements. East–West J. Numer. Math. 7, 105–131.
125
126 A. Quarteroni and L. Formaggia
F ORMAGGIA , L., N OBILE , F., Q UARTERONI , A. (2002). A one-dimensional model for blood flow: applica-
tion to vascular prosthesis. In: Babuska, I., Miyoshi, T., Ciarlet, P.G. (eds.), Mathematical Modeling and
Numerical Simulation in Continuum Mechanics. In: Lect. Notes Comput. Sci. Eng. 19 (Springer-Verlag,
Berlin), pp. 137–153.
F ORMAGGIA , L., N OBILE , F., Q UARTERONI , A., V ENEZIANI , A. (1999). Multiscale modelling of the cir-
culatory system: a preliminary analysis. Comput. Visual. Sci. 2, 75–83.
F UNG , Y.C. (1984). Biodynamics: Circulation (Springer-Verlag, New York).
F UNG , Y.C. (1993). Biomechanics: Mechanical Properties of Living Tissues (Springer-Verlag, New York).
G ASTALDI , L. (2001). A priori error estimates for the arbitrary Lagrangian Eulerian formulation with finite
elements. East–West J. Numer. Math. 9 (2), 123–156.
G IRAULT , V., R AVIART , P.-A. (1986). Finite Element Methods for Navier–Stokes Equations, Theory and
Algorithms, Springer Ser. Comput. Math. 5 (Springer-Verlag, Berlin).
G ODLEWSKI , E., R AVIART , P.-A. (1996). Numerical Approximation of Hyperbolic Systems of Conservation
Laws, Appl. Math. Sci. 118 (Springer-Verlag, New York).
G RANDMONT , C., G UIMET , V., M ADAY , Y. (2001). Numerical analysis of some decoupling techniques for
the approximation of the unsteady fluid structure interaction. Math. Models Methods Appl. Sci. 11 (8),
1349–1377.
G RANDMONT , C., M ADAY , Y. (1998). Nonconforming grids for the simulation of fluid–structure interaction.
In: Domain Decomposition Methods 10, Boulder, CO, 1997. In: Contemp. Math. 218 (Amer. Math. Soc.,
Providence, RI), pp. 262–270.
G RANDMONT , C., M ADAY , Y. (2000). Fluid structure interaction: a theoretical point of view. In: Dervieux,
A. (ed.), Revue européenne des éléments finis 9 (Hermes Science), pp. 633–653.
G UERMOND , J.-L. (1999). Un résultat de convergence d’ordre deux en temps pour l’approximation des
équations de Navier–Stokes par une technique de projection incrémentale. M2AN Math. Model. Numer.
Anal. 33 (1), 169–189.
G UILLARD , H., FARHAT , C. (2000). On the significance of the geometric conservation law for flow compu-
tations on moving meshes. Comput. Methods Appl. Mech. Engrg. 190 (11–12), 1467–1482.
H AYASHI , K., H ANDA , K., N AGASAWA , S., O KUMURA , A. (1980). Stiffness and elastic behaviour of human
intracranial and extracranial arteries. J. Biomech. 13, 175–184.
H EDSTROM , G.W. (1979). Nonreflecting boundary conditions for nonlinear hyperbolic systems. J. Comput.
Phys. 30, 222–237.
H OLZAPFEL , G.A., G ASSER , T.C., O GDEN , R.W. (2000). A new constitutive framework for arterial wall
mechanics and a comparative study of material models. J. Elasticity 61, 1–48.
H UGHES , T.J., F RANCA , L.P., BALESTRA , M. (1986). A new finite element formulation for computational
fluid dynamics: V. Circumventing the Babuska Brezzi condition: a stable Petrov–Galerkin formulation of
the Stokes problem accommodating equal-order interpolation. Comput. Methods Appl. Mech. Engrg. 59,
85–99.
L ANGEWOUTERS , G.L., W ESSELING , K.H., G OEDHARD , W.J.A. (1984). The elastic properties of 45 hu-
man thoracic and 20 abdominal aortas in vitro and the parameters of a new model. J. Biomech. 17, 425–435.
L AX , P.D. (1973). Hyperbolic Systems of Conservation Laws and the Mathematical Theory of Shock Waves,
Ser. Appl. Math. 11 (SIAM, Philadelphia, PA).
L E TALLEC , P., M OURO , J. (2001). Fluid structure interaction with large structural displacements. Comput.
Methods Appl. Mech. Engrg. 190, 3039–3067.
L IONS , J.L., M AGENES , E. (1968). Problèmes aux Limites non Homogènes et Applications, 1 (Dunod, Paris).
M EYER , C.D. (2000). Matrix Analysis and Applied Linear Algebra (SIAM, Philadelphia, PA).
M IGLIAVACCA , F., L AGANÁ , K., P ENNATI , G., DE L EVAL , M., B OVE , E., D UBINI , G. (2004). Global
mathematical modeling of the norwood circulation: a multiscale approach for the study of pulmonary and
coronary perfusions. Cardiology in the Young. In press.
N OBILE , F. (2001). Numerical approximation of fluid–structure interaction problems with application to
hemodynamics. PhD thesis, École Polytechnique Fédérale de Lausanne (EPFL), thesis N. 2458.
P EROT , B. (1993). An analysis of the fractional step method. J. Comput. Phys. 108, 51–58.
P IETRABISSA , R., Q UARTERONI , A., D UBINI , G., V ENEZIANI , A., M IGLIAVACCA , F., R AGNI , S. (2000).
From the global cardiovascular hemodynamics down to the local blood motion: preliminary applications
of a multiscale approach. In: Oñate, E., et al. (eds.), ECCOMAS 2000, Barcelona.
References 127
P IPERNO , S., FARHAT , C. (2001). Partitione procedures for the transient solution of coupled aeroelastic
problems. Part ii: energy transfer and three-dimensional applications. Comput. Methods Appl. Mech. En-
grg. 190, 3147–3170.
P ROHL , A. (1997). Projection and Quasi-compressibility Methods for Solving the Incompressible Navier–
Stokes Equations (Teubner, Stuttgart).
Q UARTERONI , A., R AGNI , S., V ENEZIANI , A. (2001). Coupling between lumped and distributed models for
blood problems. Comput. Visual. Sci. 4, 111–124.
Q UARTERONI , A., S ALERI , F., V ENEZIANI , A. (1999). Analysis of the Yosida method for the incompress-
ible Navier–Stokes equations. J. Math. Pure Appl. 78, 473–503.
Q UARTERONI , A., S ALERI , F., V ENEZIANI , A. (2000). Factorization methods for the numerical approxima-
tion of the incompressible Navier–Stokes equations. Comput. Methods Appl. Mech. Engrg. 188, 505–526.
Q UARTERONI , A., T UVERI , M., V ENEZIANI , A. (2000). Computational vascular fluid dynamics: Problems,
models and methods. Comput. Visual. Sci. 2, 163–197.
Q UARTERONI , A., VALLI , A. (1994). Numerical Approximation of Partial Differential Equations (Springer-
Verlag, Berlin).
Q UARTERONI , A., VALLI , A. (1999). Domain Decomposition Methods for Partial Differential Equations
(Oxford Univ. Press, New York).
Q UARTERONI , A., V ENEZIANI , A. (2003). Analysis of a geometrical multiscale model based on the coupling
of ODE’s and PDE’s for blood flow simulations. Multiscale Model. Simul. 1 (2), 173–195.
Q UARTERONI , A., V ENEZIANI , A., Z UNINO , P. (2002). Mathematical and numerical modelling of solute
dynamics in blood flow and arterial walls. SIAM J. Numer. Anal. 39 (5), 1488–1511.
R AJAGOPAL , K.R. (1993). Mechanics of non-Newtonian fluids. In: Galdi, G., Necas, J. (eds.), Recent Devel-
opments in Theoretical Fluid Mechanics. In: Pitman Res. Notes Math. Ser. 291 (Longman, Harlow).
R APPITSCH , G., P ERKTOLD , K. (1996). Pulsatile albumin transport in large arteries: a numerical simulation
study. ASME J. Biomech. Eng. 118, 511–519.
R EDDY , B.D. (1998). Introductory Functional Analysis. With Applications to Boundary Value Problems and
Finite Elements (Springer-Verlag, New York).
S EGEL , L.A. (1987). Mathematics Applied to Continuum Mechanics (Dover, New York).
S ERRIN , J. (1959). Mathematical principles of classical fluid mechanics. In: Flugge, S., Truesdell, C. (eds.),
Handbuch der Physik, VIII/1 (Springer-Verlag, Berlin).
S IMO , J.C., F OX , D.D. (1989). On a stress resultant geometrically exact shell model, Part I: formulation and
optimal parametrization. Comput. Methods Appl. Mech. Engrg. 72, 267–304.
S IMO , J.C., F OX , D.D., R IFAI , M.S. (1989). On a stress resultant geometrically exact shell model, Part II:
the linear theory; computational aspects. Comput. Methods Appl. Mech. Engrg. 73, 53–92.
S MITH , N., P ULLAN , A., H UNTER , P. (2003). An anatomically based model of coronary blood flow and
myocardial mechanics. SIAM J. Appl. Math. 62 (3), 990–1018.
TAYLOR , C.A., D RANEY , M.T., K U , J.P., PARKER , D., S TEELE , B.N., WANG , K., Z ARINS , C.K.
(1999). Predictive medicine: Computational techniques in therapeutic decision-making. Comput. Aided
Surgery 4 (5), 231–247.
T EMAM , R. (1984). Navier–Stokes Equations, Theory and Numerical Analysis, second ed. (North-Holland,
Amsterdam).
T HOMPSON , K.W. (1987). Time dependent boundary conditions for hyperbolic systems. J. Comput. Phys. 68,
1–24.
V ENEZIANI , A. (1998). Mathematical and numerical modelling of blood flow problems, PhD thesis, Politec-
nico di Milano, Italy.
W OMERSLEY , J.R. (1955). Method for the calculation of velocity, rate of flow and viscous drag in arteries
when the pressure gradient is known. J. Physiol. 127, 553–563.
Computational Methods
for Cardiac Electrophysiology
Mary E. Belik, Taras P. Usyk, Andrew D. McCulloch
Department of Bioengineering, University of California, San Diego, 9500 Gilman Drive,
Mail Code 0412, La Jolla, CA 92093-0412, USA
E-mail addresses: mellen@bioeng.ucsd.edu (M.E. Belik),
tusyk@ucsd.edu (T.P. Usyk),
amcculloch@bioeng.ucsd.edu (A.D. McCulloch)
Abstract
Computational methods for tissue biomechanics, electrophysiology, and cellular
physiology separately provide frameworks for modeling functions of cardiac tis-
sue. We review strategies currently available for meeting the goal of structurally
and functionally integrated models of cardiac electromechanical function that com-
bine data-intensive cellular systems models with compute-intensive anatomically
detailed multiscale simulations.
1. Background
A fundamental goal of physiology is to identify how the cellular and molecular structure
of tissues and organs gives rise to their function in vivo. Correspondingly, a key goal of
in silico physiology is to develop computational models that can predict physiologi-
cal function from quantitative measurements of tissue, cellular, or molecular structure.
Computational modeling provides a potentially powerful way to integrate structural
properties measured in vitro to physiological functions measured in vivo. It also pro-
vides a mechanism for integrating biophysical theory with experimental observation.
In this chapter, we are interested in cardiac electromechanical function, i.e., how the
cellular and extracellular organization and function of myocardial tissue is integrated
into the electromechanically coupled activation and pumping function of the whole
129
130 M.E. Belik et al.
heart. For example, how does myocardial fiber architecture influence the relation be-
tween the biophysics of action potential propagation and the three-dimensional me-
chanics of the ventricular chambers?
The physics of the heart and other organs are complex. Geometry, structure, and
boundary conditions are often irregular, three-dimensional, non-homogeneous, and time
varying. Constitutive properties and reaction kinetics are typically nonlinear and time
dependent. Fundamental physiological functions include mechanical responses and
electrical, chemical, thermal, and transport processes in cells and tissues. Therefore,
computational methods are needed to realistically model many of these diverse and
multidisciplinary processes and their integrated interactions encountered in electrophys-
iology, biomechanics, and tissue engineering.
Structural models are usually based on in vitro measurements of anatomy, tissue ar-
chitecture and material properties, and cell biophysics. Their results must be validated
with measurements from experiments conducted in vivo or in the whole isolated organ.
This iteration between model and experiment also provides the opportunity for numer-
ical hypothesis testing and in vivo constitutive parameter estimation. Once validated,
the computational models have multidisciplinary applications to problems in medicine,
surgery, and bioengineering like diagnostic imaging, surgical planning and intervention,
medical therapy, and biomedical engineering design for tissue engineering or medical
devices.
In addition to structural integration across scales of tissue organization from mus-
cle and cell to organ and system, computational models also provide a foundation for
functional integration across interacting biological processes. Computational models
have been developed for a variety of physiological processes that can be coupled for
more accurate modeling of the heart. These include biomechanics, ionic currents and
action potential propagation, contractile dynamics, energy metabolism, and cell sig-
naling. By developing a comprehensive model of cardiac electromechanics, we will
also have a framework for developing integrated models of functional interactions such
as excitation–contraction coupling, mechanoelectric feedback, mechanoenergetics, and
mechanotransduction. This development goal of integration is common and has been
considered by others in references such as H UNTER , ROBBINS and N OBLE [2002],
G IMA and RUDY [2002], KOHL , H UNTER and N OBLE [1999], N OBLE [2001], N OBLE
[2002], RUDY [2000], S UNDNES , L INES and T VEITO [2001], W INSLOW, S COLLAN ,
H OLMES , Y UNG , Z HANG and JAFRI [2000].
For many applications in cardiac physiology, the dynamic biophysical processes
within the cell and their functional interactions can be expressed by systems models
typically consisting of coupled sets of nonlinear ordinary differential equations (ODEs)
such as the common pool ionic models of myocyte electrophysiology. Similarly lumped
parameter ODE models have been developed of other biophysical processes such as en-
ergy metabolism (JAFRI , D UDYCHA and O’ROURKE [2001]) and crossbridge dynamics
(L ANDESBERG and S IDEMAN [1994]). This opens the prospect of functionally coupled
cellular models such as models of excitation–contraction coupling (M ICHAILOVA and
S PASSOV [1992], M ICHAILOVA and S PASSOV [1997]) or mechanoenergetics (TAYLOR
and S UGA [1993]) because most of the functional coupling in cardiac physiology orig-
inates within the cell.
Computational methods for cardiac electrophysiology 131
These cellular processes are spatially coupled at the tissue and organ scales, and
physico-chemical principles such as mass or momentum conservation have been used to
derive continuous field models of the resulting spatially heterogeneous behavior. These
field equations are partial differential equations that also incorporate additional dynam-
ical state variables governed by the empirical systems of ODEs described above. Thus
the ODEs describe local cellular biophysical properties and the PDEs provide a means
for structural integration from cell to organ as a three-dimensional continuum.
The geometry of the heart and other tissues and organs is complex and three-
dimensional with nonhomogeneous boundary conditions and anisotropic microstruc-
tures. The solution of coupled nonlinear PDEs and ODEs on these domains invariably
requires the use of numerical methods.
In the following section, we illustrate these concepts by deriving first some sets of
ODEs used to model the action potential in a single cardiac myocyte in terms of voltage
dependent transmembrane ionic currents, and second the reaction–diffusion equations
used to model the spread of the electrical impulse through the myocardium modeled
as a bidomain continuum. Then we introduce methods for modeling the geometry and
anatomical structure of the heart, and finally we discuss numerical methods suitable for
solving these equations and some of their present and future applications.
2. Cell biophysics
2.1. Cellular electrophysiology
The properties of the cell membrane allow for the existence of an imbalance of total
ionic charge between the intracellular and extracellular spaces. This potential difference,
the resting membrane potential, is mainly due to differing concentrations of ions across
the membrane, namely Na+ , K+ , Mg2+ , Ca2+ , H+ , and Cl− . The membrane potential
varies as ion concentrations change under various conditions and stimuli. Fig. 2.1 shows
the general shape of an action potential which occurs when a stimulus, for example an
injection of current from an electrode, causes the membrane potential to rise until a
threshold level of voltage is reached (region 1 on the schematic). This is followed by
depolarization (region 2 to 3) caused mainly by an influx of Na+ ions. In reaction,
F IG . 2.1. Schematic of an action potential.

fluxes of other ions, mainly K+ and Ca2+ , occur repolarizing the cell (region 4) which
eventually returns to a resting equilibrium state (region 5).
Ions may cross the semi-permeable cell membrane by diffusion or by active trans-
port. By Fick’s law, ions will diffuse from regions of higher concentration to regions
of lower concentration. For diffusion of charge carrying ions, a potential difference be-
tween regions may result which introduces a second driving force due to the presence of
a non-zero electrical field. The electrochemical equilibrium between the forces due to a
concentration gradient and forces due to a potential gradient for a particular ion is de-
scribed by the Nernst equation (2.1) where R, F , and T are the gas constant, Faraday’s
constant, and temperature, respectively. V is the Nernst equilibrium potential, z is the
valence of the ion in question, and Co and Ci are the concentrations outside and inside
the cell membrane.
RT [Co ]
V= ln . (2.1)
zF [Ci ]
The Goldman–Hodgkin–Katz equation (2.2) accounts for electrochemical equilibrium
of all ions present in a system; where Pion is the permeability of the membrane to a
particular ion, and Vm is the membrane potential:

RT Pion [Co ]ion
Vm = ln ion . (2.2)
zF ion Pion [Ci ]ion
Existence of electrochemical gradients across the membrane causes facilitated dif-
fusion through passive ion channels. A passive channel, when in an open state, allows
passage of a particular ion across the cell membrane with a direction and rate deter-
mined by the forces of electrochemical equilibrium. In excitable cells such as myocytes,
voltage-gating is an important determinant of the state of a channel. A closed channel at
rest will change conformation to an open state when depolarization causes the thresh-
old membrane potential to be reached. Various channels have other stable states other
than merely open or closed such as open but inactivated. Active transport is accom-
plished by ion channels that use the energy of ATP hydrolysis to transport ions against
the electrochemical gradient.
Differences in ion channel expression and resulting action potential morphology can
be found between species. In addition, action potentials vary due to cell type, which is
determined by the types of ion channels possessed and their mode of expression and
varies with the myocardial region from which the cells originate. Mammalian ventric-
ular tissue is thought to be composed of layers with different ionic properties and thus
different action potential morphologies (YAN , S HIMIZU and A NTZELEVITCH [1998]).
The ionic channels implicated in distinguishing transmural cell types are IKs , INa ,
and Ito . IKs has been found to be smaller in midmyocardial cells as compared to en-
docardial and epicardial cells, while INa and Ito are found to be larger (W OLK , C OBBE ,
H ICKS and K ANE [1999], V ISWANATHAN , S HAW and RUDY [1999], Z YGMUNT, E D -
DLESTONE , T HOMAS , N ESTERENKO and A NTZELEVITCH [2001]). In guinea pigs, it
has been found that significantly larger differences in action potential duration (APD)
exist on the epicardium from base to apex than exist between the endocardium and epi-
cardium (L AURITA , G IROUARD and ROSENBAUM [1996]). This suggests that base to
F IG . 2.2. Propagation travels from cell to cell spreading in all directions from the initial site of activation.
apex regional differences in cells may contribute more significantly to action poten-
tial heterogeneity than transmural differences in cells. In addition, it was found that an
appropriately timed premature stimulus could reverse repolarization patterns in whole
tissue even when fiber structure and propagation direction remain the same (L AURITA ,
G IROUARD and ROSENBAUM [1996]). This suggests that base to apex regional dif-
ferences in cells may contribute more significantly to action potential duration hetero-
geneity than architectural fiber and sheet arrangement. Components of IK , IKr and IKs ,
are implicated in base to apex action potential heterogeneity (C HENG , K AMIYA , L IU ,
T SUJI , T OYAMA and KODAMA [1999]). In a given cell, the character of a particular
action potential is also highly dependent on the initial state (ROSENBAUM , K APLAN ,
K ANAI , JACKSON , G ARAN , C OHEN and S ALAMA [1991]).
An action potential is a local event describing the variation of membrane potential
with time at a particular location. Depolarization at an initial site spreads in all direc-
tions to neighboring sites through action potential propagation (Fig. 2.2). In the three-
dimensional environment of the tissue, communication of action potentials from cell to
cell is affected by local tissue architecture and direct cell coupling via gap junctions.
Action potentials measured in vitro may differ from in situ due to differences in cell-to-
cell coupling conditions (A NYUKHOVSKY, S OSUNOV and ROSEN [1996]). Propagation
in whole ventricular tissue is further discussed in Section 3.
2.2. Ion channels
Early mathematical models of ion transport through cell membrane channels were de-
veloped by H ODGKIN and H UXLEY [1952] based on experimental measurements made
on the squid axon. Measurements were carried out under voltage-clamp conditions
where the membrane potential is held at a constant voltage. Voltage clamping is accom-
plished by injecting current while monitoring membrane potential. Because the voltage
is constant, no capacitive currents exist making this technique useful for isolating the
contributions of ionic currents which have an electric circuit analog of a resistive com-
ponent. According to Ohm’s law, current through a resistor is equal to the product of
conductance and the voltage drop. For the current describing the flow of a particular
ion across the cell membrane, the driving voltage drop is theoretically the difference
between the Nernst equilibrium potential of the ion and the actual membrane poten-
tial. Hence, an ionic current is described by the following equation, where gion is the
conductance of the ion, and Vion is the Nernst equilibrium potential for the ion:
Iion = gion (Vm − Vion). (2.3)
For ion currents that pass through voltage-gated channels, gion is not a constant but a
function of voltage and time. For example, a potassium channel: the protein in the cell
membrane that forms this channel is composed of four identical subunits (gates) which
each have an equal probability of being in an open or closed state (A LBERTS , B RAY,
L EWIS , R AFF , ROBERTS and WATSON [1994]). Thus,
gK = gK n4 , (2.4)
where gK is the maximal channel conductance, n is the probability that one of the gates
is open, and n4 is the probability that all four gates are open. In the original Hodgkin–
Huxley experiments the existence of four gates was found empirically. The open prob-
ability, which equals the fraction of open gates n = open/(open + closed), is calculated
by the law of mass action, which states that the change in the probability that a gate is
open with respect to time equals the difference between the rate of closed gates opening
and the rate of open gates closing:
dn
= α(1 − n) − βn, (2.5)
dt
where α and β are the rates of opening and closing, respectively. The rates, α and β,
are voltage dependent and are found by empirically fitting experimental data.
By rewriting Eq. (2.5), physical meaning can be derived in terms of α and β,
dn (n∞ − n)
= , (2.6)
dt nτ
where
α 1
n∞ = and nτ = . (2.7)
α+β α+β
The steady state solution of Eq. (2.6) is
n(t) = n∞ + (n0 − n∞ )e−t /nτ , (2.8)

where n0 is the initial probability that a gate is open. This solution is only approximate
due to the changes of α and β with voltage which render n∞ and nτ non-constant.
nτ represents the time constant that determines the rate at which n(t) approaches n∞ .
When the time constant is sufficiently small so that the steady state is reached quickly
after an initial change, n(t) is approximately n∞ .
The Hodgkin and Huxley axon model included three currents: a constant conduc-
tance chloride current, the potassium current described above, and a sodium current.
The sodium current was found to have a more complicated behavior including an inac-
tivated state as well as the open and closed states. This resulted in Eq. (2.9) for sodium
conductance with two gating variables: m for activation of the channel and h for inacti-
vation. These gating variables are governed by empirically fit equations of the form of
Eq. (2.5),
gNa = gNa m3 h. (2.9)

Advances in technology since the time of Hodgkin and Huxley have significantly
advanced knowledge of channels and their function. Modes of ion channel measure-
ment include patch clamping techniques, where small patches of membrane rather than
whole cells are studied under voltage clamp conditions. This method has the advan-
tages of allowing precise control over the ionic contents of the spaces on both sides
of the membrane as well as allowing measurements of individual ion channels. Many
channels have been and continue to be discovered, some with much more complex be-
havior. These channel models combined appropriately with models of other myocyte
functions are used to build models of whole myocytes.
2.3. Modeling cellular kinetics
2.3.1. Basic ionic models

Simulations of whole myocytes are typically derived from a statement of conservation
of current. This includes the resistive terms that represent ion channels and a capacitive
term. The cell membrane acts as a capacitor, a non-conductive dielectric, that separates
the conductive extracellular space and the conductive cytoplasm. By the definition of
capacitance,
q = Cm Vm , (2.10)
where q is charge, and Cm is membrane capacitance (farad/cm2 ).
Changes in capaci-
tive charge of the membrane over time result in a capacitive current (Fig. 2.3). By the
definition of current, the capacitive current, Ic , is
dq dVm
Ic = = Cm . (2.11)
dt dt
Thus the equation for the Hodgkin–Huxley nerve cell with sodium, potassium, and
leakage ion channels is
dVm
−Cm = INa + IK + IL
dt
= gNa (Vm − VNa ) + gK (Vm − VK ) + gL (Vm − VL ). (2.12)
A system of ordinary differential equations (ODEs) consisting of Eq. (2.12) and the
ODEs for the gating variables, n, m, and h (equations of the form of (2.5)) is solved
simultaneously for the membrane potential as it varies with time.
F IG . 2.3. Capacitive current is created by charge recruitment near the cell membrane.
F IG . 2.4. A Beeler–Reuter ionic cell model. The size of the arrows represents relative current size.
The first ionic model adapted for mammalian cardiac cells was developed by B EELER
and R EUTER [1977]. This model requires solution of eight ODEs: the membrane poten-
tial, the myoplasmic calcium concentration, and six Hodgkin–Huxley type gating vari-
ables. The gating constants are associated with four ion channels: a fast inward sodium
current, a slow inward current primarily carried by Ca2+ ions, an outward potassium
current, and a voltage dependent outward current primarily carried by potassium ions.
Fig. 2.4 shows a schematic of a Beeler–Reuter cell.

state variables = Vm , Ca2+ i , m, h, j, d, f, x ,
dVm
−Cm = INa + Is + IK + Ix − Iext . (2.13)
dt
The external current, Iext , represents an externally applied stimulation. The ODE de-
scribing the change of calcium concentration with respect to time is a function of the
slow inward current, Is . The gating variables are governed by equations of the form of
Eq. (2.5), and the twelve associated opening and closing rates are exponential functions
of voltage fitted to measured data. Refer to B EELER and R EUTER [1977] for complete
information on these equations and parameters.
2.3.2. Second generation models

Complex interacting ionic processes not limited to ion channels give rise to action po-
tentials. Second generation models incorporate intracellular processes and the resulting
effects on intracellular concentrations of various ions. Various researchers have devel-
oped models of cardiac cellular kinetics adapted for the features particular to various
species and cell types. The Luo–Rudy model (Fig. 2.5) is based on measurements from
cell voltage clamp studies in guinea pig, and uses nine ODEs (L UO and RUDY [1991],
L UO and RUDY [1994]).

state variables = Vm , Na+ i , K+ i , Ca2+ i , Ca2+ JSR , Ca2+ NSR ,

m, h, j, d, f, x . (2.14)
The six gating variables are similar to the Beeler–Reuter, Hodgkin–Huxley based, gat-
ing variables although the parameters were fit to the particular data of the guinea pig.
The membrane potential ODE consists of the sum of three main currents: sodium, potas-
sium, and calcium,
dVm
−Cm = INa,tot + IK,tot + ICa,tot . (2.14a)
dt
F IG . 2.5. Schematic of a Luo–Rudy ionic model.
However, each of these currents represents a sum of currents that relate to the particular
ion:
INa,tot = INa + 3INaCa + 3INaK + Ins,Na + INa,b + ICaNa , (2.15)

IK,tot = IK + IK1 + IKp − 2INaK + Ins,K + ICaK , (2.16)
ICa,tot = ICa − 2INaCa + Ip(Ca) + ICa,b . (2.17)
The components of the total sodium current are the fast inward sodium current, the
sodium calcium exchanger (pumping at a ratio of 3:2 sodium to calcium ions), the
sodium potassium pump (3:2 ratio of sodium to potassium), a sodium current that passes
through a non-specific calcium activated channel, a sodium background leakage current,
and the sodium component of the current passing through the L-type calcium chan-
nel. Similar sums can be written to account for total potassium and calcium currents.
Potassium currents are the time-dependent and time-independent potassium currents,
a plateau potassium current, the sodium potassium pump, and potassium currents that
flow through non-specific calcium activated channels and through the L-type calcium
channels. The calcium currents are the calcium currents through the L-type and T-type
channels, the sodium calcium exchanger, the sarcoplasmic pump current, and a calcium
background leakage current. Of these currents, some are new in the Luo–Rudy model,
others are taken from previous models with significant structural changes, and many
were taken from earlier models with only their parameters adjusted to fit the experi-
mental data used.
In addition to intracellular calcium concentration, intracellular potassium and sodium
concentrations were included. The ODEs describing these three concentrations are de-
pendent on the total current of the ion in question. The Luo–Rudy model also offers a
more detailed description of calcium handling. The sarcoplasmic reticulum (SR) is rep-
resented by three intracellular currents and two concentrations. The SR is modeled as
having two compartments: the network SR representing the longitudinal tubules where
calcium is taken up from the myoplasm by an uptake current, and the junctional SR
representing the cisternae where calcium is released back into the myoplasm. The com-
partments have a calcium transfer current between them, and each is described with a
calcium concentration. In addition, myoplasmic calcium is buffered by calmodulin and
troponin, and junctional calcium is buffered by calsequestrin.
The Luo–Rudy model is often the basis for models in other species such as the model
of Winslow et al. for dog (W INSLOW, R ICE , JAFRI , M ARBAN and O’ROURKE [1999])
and the Puglisi–Bers model for rabbit (P UGLISI and B ERS [2001]). These models ex-
tend the Luo–Rudy model by refitting the parameters that describe specific ion channel
functions, and combining or exchanging new models of particular ion channels or ionic
processes for the original Luo–Rudy kinetics in order to build species specific models.
The main extension added by the Winslow et al. model besides fitting parameters to ca-
nine experimental data concerns the calcium dynamics. A complicated L-type channel
model and a restricted subspace was added to better describe calcium induced calcium
release. The subspace of the myoplasm is postulated to lie between the junctional SR
and the cell membrane and allows close contact of L-type channels to ryanodine recep-
tors on the junctional SR membrane. It is postulated that due to geometry, the subspace
calcium concentration differs from the bulk myoplasm and is involved in triggering
SR calcium release. The Puglisi–Bers model similarly was based on the Luo–Rudy
model with the parameters rescaled for rabbit data. The main modifications in addi-
tion to rescaling were the addition of a rescaled Winslow transient outward potassium
current, the addition of a calcium activated chloride current, and modifications of the
T-type calcium (a component of ICa ) and delayed rectifier potassium (a component of
IK ) currents.
2.3.3. Other approaches

In the interest of creating integrated models across biological scales, many tissue level
models rely on phenomenological models of cellular behavior in order to have compu-
tationally tractable problem formulations. The most basic phenomenological model was
originally developed by FitzHugh and Nagumo (FHN) (F ITZ H UGH [1961], NAGUMO ,
A RIMOTO and YOSHIZAWA [1962]) for a general excitable media. It consists of only
two variables, has cubic nonlinearity, and has an on–off behavior that results in an action
potential-like depolarization and repolarization,
du
= u(u − a)(1 − u)c1 − c2 v, (2.18)
dt
dv
= bu − b dv. (2.19)
dt
While the FHN model and its modifications (ROGERS and M C C ULLOCH [1994], KO -
GAN , K ARPLUS , B ILLETT, PANG , K ARAGUEUZIAN AND K HAN [1991]) can provide
a qualitative action potential that is relatively computationally efficient, FHN models
have two major disadvantages. First, they do not accurately reproduce features of the
cardiac action potential that are of the most interest such as the rapid upstroke of de-
polarization. Second, they are not biophysically based, limiting their usefulness. For
example, the character of the action potential computed can be changed through the
various model parameters, but these cannot be related to specific biophysical mecha-
nisms so that specific channels or interventions are not possible to simulate.
On the other hand, the biophysically detailed models present two main computational
hurdles. These suffer from inherent instabilities (E NDRESEN and S KARLAND [2000])
which involve drift of ion concentrations and eventual equilibrium disruption. In addi-
tion, the growing number of variables included in these models that must be solved si-
multaneously are increasingly computationally inefficient. These significant drawbacks
have led to the development of a range of intermediate models that provide better ef-
ficiency without retaining all the known ionic details (F ENTON and K ARMA [1998],
D UCKETT and BARKLEY [2000], B ERNUS , W ILDERS , Z EMLIN , V ERSCHELDE and
PANFILOV [2002]). For example, Bernus et al. (B ERNUS , W ILDERS , Z EMLIN , V ER -
SCHELDE and PANFILOV [2002]) have reduced a Priebe–Beuckelmann human myocyte
model (P RIEBE and B EUCKELMANN [1998]) (an extension of a Luo–Rudy myocyte)
to a six variable intermediate model. This model retains the fast kinetics of the sodium
channel for accurate depolarization, but approximates other fast kinetics that do not
have large effects on action potential shape.
2.4. Sarcomere dynamics
Potentially, ionic models may be linked to models of sarcomere dynamics through con-
tractile activation and crossbridge mechanics. The basic unit of contraction is the sar-
comere. Sarcomere dynamics, namely length–tension relations, play an important role
in active force development. The length–tension relation in muscle arises from changes
in the overlap of thick and thin filaments in the sarcomere. At the level of a single my-
ocyte, contractile activation occurs shortly after depolarization of the cell as the ionic
contents shift. The time course of cytosolic calcium concentration is also central in
determining the contractile force and has been modeled using various force–calcium re-
lations. Finally, mechanisms of crossbridge mechanics, namely crossbridge recruitment
and length sensing, determine the forces developed by the sarcomere.
Long single fiber preparations have been valuable test specimens for studying the
mechanisms of skeletal muscle mechanics. The lack of these ideal test specimens is
one main reason that cardiac muscle mechanics testing is far more difficult than skele-
tal muscle testing. Moreover, under physiological conditions, cardiac muscle cannot be
stimulated to produce sustained tetanic contractions due to the absolute refractory pe-
riod of the myocyte cell membrane. Cardiac muscle also exhibits a mechanical property
analogous to the relative refractory period of excitation. After a single isometric con-
traction, some recovery time is required before another contraction of equal amplitude
can be activated (Fig. 2.6).
Unlike skeletal muscle, in which maximal active force generation occurs at a sarcom-
ere length that optimizes myofilament overlap (∼2.1 µm), the isometric twitch tension
developed by isolated cardiac muscle continues to rise with increased sarcomere length
F IG . 2.6. Isometric testing.
in the physiological range (1.6–2.4 µm). Early evidence for a descending limb of the
cardiac muscle isometric length–tension curve was found to be caused by shortening
in the central region of the isolated muscle at the expense of stretching at the damaged
ends where the specimen was tethered to the test apparatus. If muscle length is con-
trolled so that sarcomere length in the undamaged part of the muscle is indeed constant,
or if the developed tension is plotted against the instantaneous sarcomere length rather
than the muscle length, the descending limb is eliminated ( TER K EURS , R IJNSBURGER ,
VAN H EUNINGEN and NAGELSMIT [1980]). Thus, the increase with chamber volume
of end-systolic pressure and stroke work is reflected in isolated muscle as a monotonic
increase in peak isometric tension with sarcomere length. The increase in slope of the
end-systolic pressure–volume relation (ESPVR) associated with increased contractility
is mirrored by the effects of increased calcium concentration in the length–tension re-
lation. The duration as well as the tension developed in the active cardiac twitch also
increases substantially with sarcomere length.
The relation between cytosolic calcium concentration and isometric muscle tension
has mainly been investigated in muscle preparations in which the sarcolemma has been
chemically permeabilized. Because there is evidence that this chemical “skinning” al-
ters the calcium sensitivity of myofilament interaction, recent studies have also inves-
tigated myofilament calcium sensitivity in intact muscles tetanized by high frequency
stimulation in the presence of a compound such as ryanodine that opens calcium release
sites in the sarcoplasmic reticulum. Intracellular calcium concentration was estimated
using calcium-sensitive optical indicators such as fura. Myofilaments were activated in
a graded manner by microMolar concentrations of calcium, which binds to troponin C
according to a sigmoidal relation (RUEGG [1988]). Half-maximal tension in cardiac
muscle was developed at intracellular calcium concentrations of 10−6 to 10−5 M (the
[Ca]50 ) depending on factors such as species and temperature (B ERS [1991]). Hence,
relative isometric tension T0 /Tmax may be modeled using (T OZEREN [1985], H UNTER ,
M C C ULLOCH , N IELSEN and S MAILL [1988])
T0 [Ca]n
= . (2.20)
Tmax [Ca]n + [Ca]n50
The Hill coefficient (n) governs the steepness of the sigmoidal curve. A wide variety
of values have been reported but most have been in the range of 3 to 6 (K ENTISH , TER
K EURS , R ICCIARDI , B UCX and N OBLE [1986], BACKX , G AO , A ZAN -BACKX and
M ARBAN [1995]). The steepness of the isometric length–tension relation, compared
with that of skeletal muscle is due to length-dependent calcium sensitivity. That is, the
Ca50 (M), and perhaps n as well, change with sarcomere length, L.
The isotonic force–velocity relation of cardiac muscle is similar to that of skeletal
muscle, and A.V. Hill’s well-known hyperbolic relation is a good approximation except
at larger forces greater than about 85% of the isometric value (Fig. 2.7a):
V 1 − T /T0
= . (2.21)
Vmax 1 + cT /T0
The maximal (unloaded) velocity of shortening is essentially independent of pre-
load, but does change with time during the cardiac twitch and is affected by factors
that affect contractile ATPase activity and hence crossbridge cycling rates. De Tombe
and colleagues (DE T OMBE and TER K EURS [1992]) using sarcomere length-controlled
isovelocity release experiments found that viscous forces impose a significant internal
load opposing sarcomere shortening. If the isotonic shortening response is adjusted for
the confounding effects of passive viscoelasticity, the underlying crossbridge force–
velocity relation is found to be essentially linear (Fig. 2.7b).
Cardiac muscle contraction also exhibits other significant length–history-dependent
properties. An important example is “deactivation” associated with length transients.
The isometric twitch tension redeveloped following a brief length transient that disso-
ciates crossbridges reaches the original isometric value when the transient is imposed
early in the twitch before the peak tension is reached. However, following transients
F IG . 2.7. Isotonic testing: (a) the results of an isovelocity release experiment conducted during a twitch;
(b) cardiac muscle force–velocity relation corrected for viscous forces of passive cardiac muscle which reduce
shortening velocity.
applied at times after the peak twitch tension has occurred, the fraction of tension re-
developed declines progressively since the activator calcium has fallen to levels below
that necessary for all crossbridges to reattach ( TER K EURS , R IJNSBURGER , VAN H E -
UNINGEN and NAGELSMIT [1980]).
A number of models of active tension development in cardiac muscle have been pro-
posed. They may be grouped into three categories:
(1) time-varying elastance models include the essential dependence of cardiac active
force development on muscle length and time (A RTS , R ENEMAN and V EEN -
STRA [1979], C HADWICK [1982], TABER [1991]),
(2) “Hill” models, in which the active fiber stress development is modified by short-
ening or lengthening according to the force–velocity relation, so that fiber tension
is reduced by increased shortening velocity (A RTS , V EENSTRA and R ENEMAN
[1982], N EVO and L ANIR [1989]), and
(3) fully history-dependent models that are usually based on A.F. Huxley’s cross-
bridge theory (PANERAI [1980], L ANDESBERG and S IDEMAN [1994], L AN -
DESBERG , M ARKHASIN , B EYAR and S IDEMAN [1996]) which yields a sys-
tem of partial differential equations as functions of time and crossbridge posi-
tion.
Many of the early models were based on skeletal muscle models of Hill (H ILL [1938],
H ILL [1970]). However, Hill’s model considers tetanic contraction only and hence
is inappropriate for describing cardiac muscle mechanics (F UNG [1981]). According
to PANERAI [1980], Wong (W ONG [1971], W ONG [1972]) was the first to employ
the sliding filament theory to model the mechanics of cardiac muscle. Wong gener-
alized Huxley’s model (H UXLEY [1957]) of the skeletal muscle crossbridge to partial
and length-dependent activation. Panerai, using Huxley’s original model, incorporated
length-dependent activation in a first order kinetic equation describing Ca2+ -troponin C
interaction. Instead of considering individual myofilaments, T OZEREN [1985] proposed
a “continuum” model of cardiac muscle contraction. Tozeren generalized Hill’s equation
to partial activation to describe active fiber tension as a function of fiber strain, strain
rate and time after onset of contraction. In these studies, model predictions were val-
idated by experimental length–tension or force–velocity relations during contractions
in which overall muscle length was controlled. Panerai accounted for the appreciable
internal shortening that occurs during isometric contractions at the expense of lengthen-
ing in the damaged muscle at the clamped ends (K RUGER and P OLLACK [1975], TER
K EURS , R IJNSBURGER , VAN H EUNINGEN and NAGELSMIT [1980]).
Continuum models typically compute the active tension developed by a cardiac mus-
cle fiber from the peak intracellular Ca2+ ion concentration, the time after onset of
contraction and the sarcomere length–history (G UCCIONE and M C C ULLOCH [1993]).
Model contraction is driven by a free calcium transient that is independent of length.
The number of actin sites available to react with myosin is determined from the total
number of actin sites (available and inhibited), the free calcium, the length–history-
dependent association and dissociation rates of Ca2+ ions to troponin binding sites and
the troponin concentration.
H UNTER , M C C ULLOCH and TER K EURS [1998] described intracellular calcium

concentration as a function of time (Cai (t)) as follows :
t 1−t /τCa
Cai (t) = Ca0 + (Camax − Ca0 ) e , (2.22)
τCa
where Ca0 is the resting concentration of intracellular calcium and Camax is the peak
intracellular calcium concentration which occurs at t = τCa .
Panerai described Ca2+ –troponin C interaction using a rate equation derived from
classical chemical kinetics. This equation determines the concentration of actin that is
free to react with myosin (Ac (t)) from the kinetics of the binding of two calcium ions
to independent sites on the troponin molecule,
dAc (t)
= c1 Ca2i (t) Aco − Ac (t) − c2 Ac (t), (2.23)
dt
where Aco is a constant reflecting the total amount of actin present in the muscle, and
c1 and c2 are the association and dissociation rate constants, respectively.
Ca2+ –troponin C binding results in tropomyosin movement, which in turn controls
the availability of actin binding sites. For modeling tropomyosin kinetics, H UNTER ,
M C C ULLOCH and TER K EURS [1998] introduced a non-dimensional parameter z, (0
z 1), to represent the proportion of actin sites available for crossbridge binding. They
proposed a model for z reflecting an exponential increase in tension with a first order
rate constant dependent on calcium activation,

dz Ca n
= α0 (1 − z) − z , (2.24)
dt Ca50
where α0 is the rate constant of tropomyosin movement. Ca50 and n are Hill parameters
determined for a particular sarcomere length by fitting the equilibrium relation between
z and [Ca]. Under an assumption of proportionality between steady state force and
available actin binding sites, a sigmoidal response curve (Hill-type equation) is used
to describe the steady state developed tension as a function of calcium concentration
(H UNTER , M C C ULLOCH and TER K EURS [1998]),
[Ca]n
zss = (2.25)
[Ca]n + [Ca]n50
where zss takes a value between 0 and 1, [Ca]50 is the calcium concentration required
to produce 50% of peak contraction, and n is the Hill parameter describing the shape of
the sigmoidal curve. Note that Eq. (2.25) is also known as a “Hill” equation, not to be
confused with Hill’s equation for the force–velocity relation (2.21).
Sarcomere length dependence is included by describing developed maximum ten-
sion as a function of λ, the extension ratio of sarcomeres (l0 / lref , where lref is resting
sarcomere length):

T0 = Tref 1 + β0 (λ − 1) . (2.26)
From experimental observations in rats, Tref = 125 kPa (the reference tension when
λ = 1) (H UNTER , M C C ULLOCH and TER K EURS [1998]). The term β0 describes my-
ofilament cooperativity,
1 dT0
β0 = = 1.45. (2.27)
Tref dλ
For less than full activation, the dependence on calcium under isometric conditions may
be approximated by

T0 = Tref 1 + β0 (λ − 1) zss . (2.28)
Length-dependence of the parameters n and [Ca]50 is approximated in a similar form:

n = nref 1 + β1 (λ − 1) , (2.29)

pC50 = pC50ref 1 + β2 (λ − 1) , (2.30)
[Ca]50 = 106−pC50 in µM. (2.31)
In order to fit the experimental data that K ENTISH , TER K EURS , R ICCIARDI , B UCX and
N OBLE [1986] obtained from skinned rat right ventricular muscle, the parameters of the
model were chosen as: nref = 4.25; β0 = 1.45; β1 = 1.95; β2 = 0.31; pC50ref = 5.33;
Tref = 125 kPa; Camax = 4.3 µM, the maximal intracellular calcium concentration.
Sarcomere length can affect generated force through crossbridge recruitment and
through crossbridge distortion. Recruitment affects generated force by altering the num-
bers of attached crossbridges. Although this is the basis for Starling’s law, the mecha-
nisms are not yet clearly understood. Distortion is internal stretch of a crossbridge struc-
ture. Again mechanisms for sensing stretch are not fully understood. If crossbridges are
modeled as having constant, linear elastic material properties, the total force generated
in one half of a myosin and actin filament pair is as follows:

l0
Ftotal = K N, (2.32)
lref
where K is a spring constant describing stretch of a single crossbridge, l0 / lref is the
stretch of the crossbridge, and N is the number of attached crossbridges. Cooperativity
mechanisms describe possible ways that crossbridge interactions sense and modulate
force. Examples include sensing based on lateral spacing of actin and myosin, mem-
ory based mechanisms where recruitment depends on initial length, or existence of a
crossbridge state that is attached but not yet generating force thereby accounting for ex-
ceedingly rapid recruitment rates that fall off during late stages of contraction (H UNTER
[1999], R ICE , W INSLOW and H UNTER [1999]).
3. Impulse propagation dynamics
3.1. Physiology of propagation
The nervous system directs the behavior of the heart; however, cardiac tissues have in-
trinsic rhythmicity (or automaticity) meaning cardiac tissues can initiate beats without
nervous stimulus. This pacemaking activity may serve as a safety mechanism and is
sufficient for successful cardiac function, for example, in the case of completely den-
ervated mammalian hearts used for cardiac transplant. The cardiac conduction system
in mammals begins with the sinoatrial node (SA node) located where the superior vena
cava joins the right atria. The SA node may be referred to as the natural pacemaker
of the heart, and it normally generates impulses at a higher frequency than other car-
diac tissues. A cardiac impulse is transmitted via various internodal pathways from the
SA node, across the atria to the atrioventricular node (AV node). The AV node is also
capable of producing pacemaking impulses for the whole heart, at a somewhat lower
frequency than the SA node. The impulse continues along the fibers of the AV node as
they transition into fibers belonging to the bundle of His. In normal conditions this is
the only pathway by which an impulse reaches the ventricles (see B ERNE and L EVY
[1997], K ATZ [2001]).
The bundle of His represents the upper portion of the ventricular conduction system.
It divides into left and right bundle branches one leading subendocardially to each of the
ventricles. The bundle branches continue to subdivide into the complex Purkinje fiber
network. The Purkinje fibers also possess automaticity but can only generate impulses
at a significantly slow rate. Impulse conduction through the Purkinje fibers is the fastest
of any cardiac tissue enabling rapid spread of the impulse throughout the ventricular
endocardium. The Purkinje network has been mapped in various species. In the sheep,
Purkinje fibers were found to be surrounded by a perifascicular sheath which may help
to direct conduction along the fiber network rather than into the surrounding tissue or
may provide protection from friction during contraction (A NSARI , H O and A NDERSON
[1999]). In the rabbit, while the left ventricular Purkinje system follows a branching
pattern similar to that in the dog, the right ventricle was found to have a denser web-like
arrangement (C ATES , S MITH , I DEKER and P OLLARD [2001]). The connections to the
myocardium at the ends of the Purkinje network consist of transitional cells that have
also been characterized (C ATES , S MITH , I DEKER and P OLLARD [2001], T RANUM -
J ENSEN , W ILDE , V ERMEULEN and JANSE [1991]). An impulse first excites the septum
as it spreads along the Purkinje network along the endocardium. From the ends of the
Purkinje fibers, the excitation in the ventricular walls spreads through the myocardial
tissue from endocardium to epicardium at a slower rate than the conduction through
the Purkinje system. Apical and central regions of the ventricular free walls are excited
first and activation proceeds toward the base. Spread of activation within the ventricular
myocardium itself is affected by local tissue characteristics.
The microstructure of ventricular architecture is important in determining three-
dimensional conduction patterns. As discussed in Section 5.4, ventricular myocardium
is arranged in fibers which are in turn arranged in laminar sheets (L E G RICE , S MAILL ,
C HAI , E DGAR , G AVIN and H UNTER [1995]). This structural arrangement in addition
to its affects on cardiac mechanics influences propagation. In pacing studies where stim-
ulating electrodes were introduced into the ventricular wall in order to elicit activation
at varying wall depths, clear anisotropic propagation was seen. Patterns of depolariza-
tion were found to align with the fiber architecture as demonstrated by helical rota-
tion of the pattern as stimulation varied with depth (E FIMOV, E RMENTROUT, H UANG
and S ALAMA [1996]). Various modeling efforts have also confirmed the importance
of fiber architecture to electrical propagation (F RANZONE , G UERRI , P ENNACCHIO
and TACCARDI [1998], F RANZONE , G UERRI , P ENNACCHIO and TACCARDI [2000],

M UZIKANT and H ENRIQUEZ [1997], M UZIKANT and H ENRIQUEZ [1998], K EENER
and PANFILOV [1997], S AXBERG , G RUMBACH and C OHEN [1985], ROGERS and M C -
C ULLOCH [1994], V ETTER and M C C ULLOCH [2001]). Measurements indicating de-
pendence on sheet architecture are virtually nonexistent. This is most likely due to the
difficulties associated with obtaining three-dimensional recordings, particularly inside
the ventricular wall. Conduction is known to be faster in the fiber direction compared to
the sheet and sheet normal directions (K ANAI and S ALAMA [1995]), but little is known
experimentally about the ratio of sheet to sheet normal conduction. A recent study of
the effects of tissue microstructure on propagation supports the reasonability of model-
ing conduction as orthotropic rather than transversely isotropic (H OOKS , T OMLINSON ,
M ARSDEN , L E G RICE , S MAILL , P ULLAN and H UNTER [2002]).
In addition to structural heterogeneities, functional heterogeneities exist regionally
within ventricular tissue. Studies show that the myocytes that compose endocardial
and epicardial layers of the ventricle walls have different electrical properties from
those in the mid-wall (A NTZELEVITCH , S HIMIZU , YAN , S ICOURI , W EISSENBURGER ,
N ESTERENKO , B URASHNIKOV, D I D IEGO , S AFFITZ and T HOMAS [1999], S ICOURI
and A NTZELEVITCH [1995]). Changes in ion concentrations have been shown to result
in differential regional effects supporting the variation of cell type as the underlying
mechanism (W OLK , K ANE , C OBBE and H ICKS [1998]). These midmyocardial cells
(M cells) are characterized by the ability to disproportionately prolong action poten-
tial duration compared to other myocardial cells in response to slowing of stimulation
rate or APD prolonging agents (A NTZELEVITCH , S HIMIZU , YAN , S ICOURI , W EIS -
SENBURGER , N ESTERENKO , B URASHNIKOV, D I D IEGO , S AFFITZ and T HOMAS
[1999]). This arrangement may be the basis for the T-wave morphology of electrocar-
diograms. The T-wave characterizes repolarization. The peak of the T-wave has been
found to correspond with epicardial repolarization and the end with M cell repolariza-
tion (YAN , S HIMIZU and A NTZELEVITCH [1998]). Very recent evidence suggests that
the arrangement of cell types may not consist of transmural layers. Instead M cells may
be located in islands in various regions of the ventricle wall (A KAR , YAN , A NTZELE -
VITCH and ROSENBAUM [2002]).
In the whole ventricle repolarization depends on tissue heterogeneity (W OLK ,
C OBBE , H ICKS and K ANE [1999]). Endocardial myocytes depolarize before epicardial
cells during normal sinus rhythm, while repolarization has been measured in the canine
to occur first in the epicardium followed by the endocardium and finally the midmy-
ocardium (YAN , S HIMIZU and A NTZELEVITCH [1998]). Repolarization patterns have
been found to be much more sensitive to test conditions than depolarization patterns in
transmural optical mapping studies in the canine (A KAR , YAN , A NTZELEVITCH and
ROSENBAUM [2002]). Evidence for the significance of regional heterogeneities of cell
types versus architectural fiber orientation in determining repolarization sequence seems
to be conflicting between studies in dog and guinea pig. In guinea pig repolarization
seemed to spread from apex to base regardless of pacing site while in dog repolarization
was highly dependent on activation sequence (K ANAI and S ALAMA [1995]). Corrob-
orating these findings, timed premature stimuli have been found to reverse the pattern
of repolarization even though fiber structure and propagation direction remained the
same in the guinea pig (L AURITA , G IROUARD and ROSENBAUM [1996]). Comparison
of measurements of repolarization in canine and guinea pig have led to the conclusion
that the smaller size of the guinea pig heart requires less time for complete depolar-
ization so that intrinsic regional heterogeneities dominate the pattern of repolarization
(K ANAI and S ALAMA [1995], E FIMOV, E RMENTROUT, H UANG and S ALAMA [1996],
L AURITA , G IROUARD and ROSENBAUM [1996]).
Regional variations in action potential morphology and ion channel expression have
also been observed from base to apex. In rabbit myocytes, the density of IKr and IKs ion
channels varied in basal and apical samples. IKr was expressed in higher density in the
apex than the base. In addition, the ratio of IKs to IKr was larger in the base than in the
apex. Action potential duration was found to be longer in the apex than base (C HENG ,
K AMIYA , L IU , T SUJI , T OYAMA and KODAMA [1999]). This study did not rule out the
possibility of detecting transmural variations in the apical or basal sections; however,
other observations of whole ventricle repolarization patterns also suggest base to apex
variation of cell types. The actual three-dimensional variation of cell types remains
debatable, and other regional variations are possible that have not yet been examined.
For example, typically the left and right ventricular walls have been treated as consisting
of similar cell type distributions; however, it is possible and perhaps likely that regional
variations also exist in this case (W OLK , C OBBE , H ICKS and K ANE [1999]).
3.2. Cable theory
Cable theory was developed for power transmission through uniform conducting cables
surrounded by an insulating medium and was first applied to cardiac tissue by W EID -
MANN [1970]. A muscle fiber has a conductive interior consisting of the myoplasm and
is surrounded by the insulating cell membrane. The myoplasm has a material property,
D, that describes its three-dimensional conductivity (mS/cm).
An electric field vector (mV/cm), E, is defined as a potential drop maintained spa-
tially in a material,
E = −∇Φi , (3.1)
where
∂ ∂ ˆ ∂
∇= î + j+ k̂ (3.2)
∂x1 ∂x2 ∂x3
and Φi is the scalar function for potential inside the cable (or the intracellular potential,
hence the subscript i). By Ohm’s law, the flux vector (µA/cm2 ), J, which represents the
current density inside the cable, is proportional to the electric field vector,
J = DE = −D∇Φi . (3.3)
Physically, Eq. (3.3) states that current flux in a cable occurs in the direction of the
greatest potential drop.
Considering a differentially small section of cable with volume, dΩ, and surface
area, dΓ = dΓe + dΓm , the total current that enters the section of cable must equal the
current that leaves as dictated by the conservation of charge (see Fig. 3.1). Since charge
F IG . 3.1. Current in a section of cable. Inward currents are positive, and outward currents are negative.
cannot accumulate within the cable, the flux in through area dΓe minus flux out through
dΓe (fluxes through the cable ends to neighboring conductive volumes) and the sum
of membrane currents (per unit area) that cross the membrane through dΓm (positive
membrane currents defined as flowing from inside to out) must balance:

J − (J + dJ) dΓe − (Ic + Iion ) dΓm = 0, (3.4)
where the total differential of flux over the length of cable, dJ , in three dimensions is
∂Jx1 ∂Jx2 ∂Jx3
dJ = dx1 + dx2 + dx3 , (3.5)
∂x1 ∂x2 ∂x3
and
dVm
Ic = Cm , (3.6)
dt
where Ic is the capacitive membrane current. Iion is the sum of the currents that cross the
membrane through ion channels with units (µA/cm2 ). Then the units of the right-hand
side of Eq. (3.4) are (µA/cm2 )(cm2 ) = µA.
Rearranging Eq. (3.4),
−(dJ) dΓe = (Ic + Iion ) dΓm . (3.7)
Examining the terms of the right-hand side, the change in flux in the x1 direction, for
example, is

∂Jx1 ∂Jx1
− dx1 dx2 dx3 = − dx1 dx2 dx3, (3.8)
∂x1 ∂x1
where dx2 dx3 = dΓe , and dx1 dx2 dx3 = dΩ. In the general 3D case, the total change
in flux is

∂Jx1 ∂Jx2 ∂Jx3
−(dJ) dΓe = − + + dΩ = −(∇ · J) dΩ. (3.9)
∂x1 ∂x2 ∂x3
Substituting Eq. (3.9) into (3.7) results in the following statement of the conservation of
charge:

dVm
−(∇ · J) dΩ = Cm + Iion dΓm . (3.10)
dt
By setting dΓm /dΩ = Sv , the cell membrane surface area to volume ratio, and express-
ing flux in terms of potential from Eq. (3.3), Eq. (3.10) results in the three-dimensional
cable equation:
1 dVm 1
∇ · D∇Φi = + Iion . (3.11)
Sv Cm dt Cm
It should also be noted that the conductivity tensor, D, is often written with units of
cm2 /msec by combining D with Sv and Cm . Thus the cable equation may also be written
as
dVm 1
∇ · D∇Φi = + Iion . (3.12)
dt Cm
The cable equation can also be derived by considering integral forms of the statement
of charge conservation rather than the differential form presented above. Considering
an arbitrary volume of cable, Ω, with surface, Γ = Γe + Γm , and outward normal n, an
equivalent statement of Eq. (3.4) is

dVm
n · J dΓ + Cm + Iion dΓ = 0, (3.13)
Γe Γm dt
where outward fluxes and currents are positive. Applying the Gauss theorem directly
results in

∂Ji
n · J dΓ = dΩ = ∇ · J dΩ, (3.14)
Γ Ω ∂xi Ω
and Eq. (3.13) becomes

dVm
∇ · J dΩ = − Cm + Iion dΓ. (3.15)
Ω Γm dt
Since this must hold for any arbitrary volume and associated surface area of the cable

dVm
∇ · J = −Sv Cm + Iion , (3.16)
dt
where Sv accounts for the fact that the membrane currents flow only through a propor-
tionally sized portion of the surface area of the arbitrary volume. Applying the results
of Ohm’s law from Eq. (3.3), Eq. (3.16) also leads to the cable equation (Eq. (3.11)).
3.3. Governing equations of cardiac propagation
The bidomain model for describing the electrophysiology of cardiac tissue was intro-
duced by G ESELOWITZ and M ILLER 3 RD [1983]. This model treats cardiac tissue as
two separate, continuous domains that both occupy the space occupied by the tissue.
These two domains represent the intracellular and extracellular spaces. ui and ue are
the electrical potentials in each domain. According to the conservation of current, any
communication of current between the two domains is considered to have crossed the
cell membrane. Current flowing strictly within either domain can be assumed to be
purely resistive since in the actual microstructure of intra- and extracellular domains,
that occupy mutually exclusive regions of space, continuous pathways exist. In the ex-
tracellular space current can flow continuously between the cells, and in the intracellular
space current can flow continuously through cells since their myoplasm is connected via
gap junctions. The three-dimensional resistance to current flow in each domain is de-
scribed by a tensor of conductivity parameters.
Applying Ohm’s law, Eq. (3.3), to these two domains results in
Ji = −Di ∇ui , (3.17)
Je = −De ∇ue , (3.18)
where Ji and Je are the intra- and extracellular current densities, and Di and De are
anisotropic conductivity tensors. If an arbitrary volume domain of the tissue, Ω, with
a surface Γ is defined, conservation of charge requires that the flux out of one domain
must equal the flux into the other,

Ji · n dΓ = − Je · n dΓ, (3.19)
Γ Γ
where n represents the unit outward normal to the arbitrary volume. Approximating the
behavior of the intracellular domain as a cable, as in the above derivation, application
of the Gauss theorem leads to

dVm
−∇ · Ji = Sv Cm + Iion , (3.20)
dt

dVm
−∇ · Je = −Sv Cm + Iion , (3.21)
dt
where the opposite signs indicate that the same membrane currents flow out of the ex-
tracellular domain and into the intracellular domain.
The membrane potential is defined as the difference between the intra- and extracel-
lular potentials,
V m = ui − ue . (3.22)
Substitution of Eqs. (3.17) and (3.18) into Eqs. (3.20) and (3.21) results in

dVm
∇ · Di ∇ui = Sv Cm + Iion , (3.23)
dt

dVm
∇ · De ∇ue = −Sv Cm + Iion . (3.24)
dt
Writing ui in terms of Vm and ue ,

dVm
∇ · Di (∇Vm + ∇ue ) = Sv Cm + Iion , (3.25)
dt

dVm
∇ · De ∇ue = −Sv Cm + Iion . (3.26)
dt
By writing Eq. (3.26) in terms of Eq. (3.25), Eq. (3.26) becomes
∇ · De ∇ue = −∇ · Di (∇Vm + ∇ue ). (3.27)
Finally, rearranging both equations:

dVm
∇ · (Di ∇Vm ) + ∇ · (Di ∇ue ) = Sv Cm + Iion , (3.28)
dt

∇ · (Di + De )∇ue = −∇ · (Di ∇Vm ). (3.29)
These are the governing equations for the bidomain treatment of ventricular tissue. The
unknowns Vm and ue vary spatially within the domain of the problem, in this case
the ventricles, and are thus partial differential equations. Iion , a function of Vm , is the
collection of membrane currents detailed in Section 2.3 that describe the local ionic
state of the cells in the tissue as a function of time through a set of ordinary differential
equations.
The boundary conditions needed to complete the formulation describe the situation
that on the surface of the heart, the flux in Eq. (3.21) from the extracellular space to the
intracellular space does not exist since the intracellular space does not exist outside of
the heart. Eq. (3.29) is an equivalent statement to (3.21), so the boundary condition can
be derived as

∇ · (Di + De )∇ue = −∇ · (Di ∇Vm ) = 0 on Γ, (3.30)
or
n · (Di ∇Vm ) = 0 on Γ. (3.31)
Finite elements can be used to solve these governing equations with the associated
boundary conditions. We return to the solution of the bidomain field equations for car-
diac electrophysiology after a derivation of some applicable finite element methods in
Section 4.
4. Finite element methods
The finite element method is a popular computational approach to these problems that
has applications in diverse areas of cardiovascular biophysics, such as
• blood flow in arteries;
• stress and strain distributions in the myocardium of the beating heart;
• bioheat transfer in myocardial tissue during laser, cryo or radio-frequency ablation;
• multicellular action potential propagation;
• shock and defibrillation;
• strain analysis from cardiac magnetic resonance imaging;
• the inverse electrocardiographic problem.
In this section we derive finite element equations useful for both constructing models
of complex domains such as the heart and for solving the bidomain equations for action
potential propagation. More detailed texts on finite element methods can be found in
C IARLET and L IONS [1990], C AREY and O DEN [1983], L ANGTANGEN [1999], L I -
ONS and M AGENES [1972] and elsewhere. Results from this section are applied to the
construction of models of cardiac geometry and structure that provide the domain on
which the bidomain equations may be solved in Section 5. Then in Section 6, results
from this section are applied to the bidomain equations themselves.
4.1. Formulation of FE equations
In the general case, an unknown function u satisfies a certain partial differential equa-
tion (PDE) represented by the differential operator L such that L(u) is satisfied on a
domain, Ω,
L(u) = 0 on Ω, (4.1)
subject to appropriate boundary conditions on the boundaries, Γ ,
B(u) = 0 on Γ. (4.2)
In general, boundaries can be divided into two non-intersecting sections: one where
conditions are prescribed on the derivative of the solution, u, and one where conditions
are prescribed on u itself. In the following discussion, the boundary, Γ , is the portion
of the boundary where conditions are prescribed on the derivative of u because all ap-
proximations to the solution are always chosen such that prescribed conditions on u are
explicitly satisfied.
Eqs. (4.1) and (4.2) are the strong formulation since they require an exact solution u
everywhere. The finite element method seeks an approximate solution û in the form

M
u û = Ψi u i , (4.3)
i=1
where Ψ are basis functions prescribed in terms of independent variables (such as spa-
tial coordinates x, y, z), and some or all of the coefficients ui are unknown. In deter-
mining u the goal is to minimize the error between û and u.
In general, because û is not the exact solution, substitution of the approximation into
the differential equation results in
L(û) = 0, (4.4)
B(û) = 0. (4.5)
Since the exact solution is generally unknown, the differences between L(u) and L(û)
and B(u) and B(û) are used as a measure of error. By definition, L(u) and B(u) equal
zero, so L(û) and B(û) become the measure of error, the residual,
L(û) + B(û) = R. (4.6)
Various procedures for minimizing R give rise to finite element methods including
weighted residuals, collocation, and least squares.
Because the residual, R, is defined over a domain, Ω, it varies over the independent
variables (for the present problem, it varies spatially over the domain of the ventricles).
In order to find a minimum, a scheme for integrating R over the domain will be used to
find an average measure of the error in the approximation L(û) and B(û) due to û = u.
Since the PDE in Eq. (4.1) must be zero at each point in the domain, Ω, it follows that

ωL(u) dΩ ≡ 0, (4.7)
Ω
where ω is an arbitrary weighting function. The same integral equation can be written
for the boundary conditions

ωB(u) dΓ ≡ 0, (4.8)
Γ
for any function ω. The integral statement that

ωL(u) dΩ + ωB(u) dΓ = 0, (4.9)
Ω Γ
is satisfied for all ω and ω and is equivalent to the PDE in Eq. (4.1) and the boundary
conditions in Eq. (4.2).
If the unknown function is approximated by the expansion in Eq. (4.3), the result is
the weighted residual formulation,

ωi L(û) dΩ + ωj B(û) dΓ = 0,
Ω Γ
where i = 1, . . . , m, j = m + 1, . . . , M. (4.10)
m is the number of unknowns, ui , and M − m is the number of unknowns on the bound-
ary. In solving Eq. (4.10) for the approximate solution û, the PDE is satisfied only in
an average sense, so the weighted residual is a weak formulation. Certain choices of the
weight functions result in the point collocation and least squares methods.
4.1.1. Point collocation method

Using the Dirac delta function (which equals x when x = c, but equals zero otherwise)
in choosing the weight functions results in the collocation method:

m
ω(x) = δ(x − ci ), (4.11)
i=1

M
ω(x) = ρ 2 δ(x − ci ). (4.12)
i=m+1
The constant ρ is introduced for consistency of units. The following property of the
Dirac delta function

f (x)δ(x − ci ) dΩ = f (ci ), (4.13)
Ω
results in

L û(ci ) = 0 for i = 1, . . . , m on Ω, (4.14)

B û(ci ) = 0 for i = m + 1, . . . , M on Γ, (4.15)
when these weight functions are applied to the weighted residual equation (Eq. (4.10)).
The approximate solution is forced to satisfy the PDE only at M points in the body and
on the boundary where the dependent variable x is equal to ci .
4.1.2. Least squares method

The least squares method arises from seeking the minimum of the averaged square of
the residual. The averaged square of the residual is

L(û) dΩ + ρ
2 2
B(û)2 dΓ. (4.16)
Ω Γ
The constant ρ is again introduced for consistency of units. Take the first variation to
find the minimum

∂
L(û) dΩ + ρ
2 2
B(û) dΓ = 0,
2
(4.17)
∂ui Ω Γ

∂L(û) ∂B(û)
2L(û) dΩ + ρ 2 2B(û) dΓ = 0, where i = 1, . . . , M. (4.18)
Ω ∂ui Γ ∂ui
This is the same as choosing the weight functions as follows:
∂L(û)
ω(x) = 2 where i = 1, . . . , m on Ω, (4.19)
∂ui
∂B(û)
ω(x) = 2ρ 2 where i = m + 1, . . . , M on Γ. (4.20)
∂ui
4.2. Boundary conditions
The following generic boundary value problem can be used to illustrate the treatment
of boundary conditions:

−∇ · k(X)∇u(X) = f (X) on Ω, (4.21)
−n · k(X)∇u(X) = g(X) on ΓN , (4.22)
u(X) = φ(X) on ΓE , (4.23)
where the boundary has been divided into two sections. It is important to note that this
model has the same form as the bidomain equations and its boundary conditions, so the
result here is directly applicable and will be described in Section 6.
A term added to the approximation in Eq. (4.3) with a requirement that Ψi = 0 on ΓE
causes the essential boundary conditions prescribed on ΓE to be satisfied automatically
by the choice of the following form for û:

M
û = φ + Ψi u i . (4.24)
i=1
Applying the weighted residual formulation of Eq. (4.10):

− ∇ · [k∇ û] + f ωi dΩ − (n · k∇ û + g)ωi dΓ = 0,
Ω ΓN
where i = 1, . . . , M (4.25)
for N linearly independent weight functions. Integrating the first term in the first inte-
grand by parts:

∇ωi · k∇ û dΩ − (n · k∇ û)ωi dΓ − f ωi dΩ
Ω ΓN Ω

− (n · k∇ û + g)ωi dΓ = 0. (4.26)
ΓN
If the weight functions are chosen so that ωi = Ψi and ωi = −Ψi (the Galerkin pro-
cedure), then the terms concerning the boundary conditions prescribed on the natural
boundary “naturally” cancel from the equation, leaving

∇ωi · k∇ û dΩ − f ωi dΩ − gωi dΓ = 0. (4.27)
Ω Ω ΓN
Inserting the approximate solution, û, and the chosen weighting functions:
M

∇Ψi · k∇Ψj dΩ uj = f Ψi dΩ − gΨi dΓ
j =1 Ω Ω ΓN
where i = 1, . . . , M. (4.28)
This yields a set of M linear equations in the form
Ku = f, (4.29)
where K and f are known and u = ui can be found using a linear solver.
4.3. Domain discretization
Because the problem of finding an approximation of u to satisfy L(u) has been ex-
pressed in an integral form over the domain Ω in Eq. (4.9), the properties of integration
can be taken advantage of in order to make a continuous, intractable problem tractable
with a finite number of unknowns. The key property is that of summation: an integral
over an arbitrary domain is the same as the sum of integrals over a set of arbitrary
non-overlapping subdomains whose union is the original domain. Domain discretiza-
tion permits the approximation to be obtained element by element and an assembly to
be achieved by summation.
In the case of the finite element method, the approximate solution, û, becomes a set
of piece-wise functions, û, spanning separate subdomains (the elements),
n
max
rmax
u û = r
Ψn(e) urn(e) , (4.30)
n(e)=1 r=1
where M = nmax × γmax terms make up each component of û which has one component
for each element. Recasting the original PDE in Eqs. (4.1) and (4.2) into subdomains
over a domain Ω, subject to appropriate boundary conditions on the boundaries Γ :

 

 A1 (u) 

A(u) = A2 (u) = 0, (4.31)

 ..  
.
 

 B1 (u) 

B(u) = B2 (u) = 0, (4.32)

 ..  
.
where A is a set of piece-wise PDE operators on the subdomains with B on the boundary
subdomains equivalent to the original PDE, L, on the whole domain with B on the
boundaries. The weak formulation can then be written in the following form:
max
e
Fj (û) dΩ + fj (û) dΓ = Fj (û) dΩ +
e e
fj (û) dΓ , (4.33)
Ω Γ e=1 Ωe Γe
where Ω e is the domain of each element and Γ e is its part of the boundary. Fj and fj
prescribe known functions or operators.
Considering linear partial differential equations and boundary conditions:
A(u) = Lu − f = 0 in Ω, (4.34)
B(u) = Mu − p = 0 on Γ. (4.35)
The approximation scheme (4.33) will again yield a set of linear equations in the form
Ku + f = 0. (4.36)
4.4. Two uses of basis functions
The set of basis functions, Ψ , is generally used for two purposes: the approximation of
the unknown dependent variable (as in Eq. (4.3)) and the parameterization of the de-
pendent variables defining the problem domain into the finite elements. In the case that
the same set of basis functions is used for both purposes, the result is an isoparametric
element interpolation with an isoparametric mapping of element to global coordinates,
i.e., an isoparametric interpolation for a field variable, u(x), is created by defining the
geometric (dependent) variable x as an interpolation of nodal parameters using the same
basis functions:

u(ξ ) = M i=1 Ψi (ξ )ui
⇒ u = u(x). (4.37)
x(ξ ) = M i=1 Ψ i (ξ )x i
4.4.1. Dependent variable interpolation

To approximate a set of points {xk ; uk (xk )} by a continuous function, a convenient and
popular method is to use a polynomial expression such as: u(x) = a + bx + cx 2 +
dx 3 + · · · and then to estimate the monomial coefficients a, b, c, d, . . . to obtain a best
approximation to the field variable u(x). Since high-order polynomials, such as quartics
F IG . 4.1. Parameterizing piece-wise polynomials in terms of shared nodal parameters automatically ensures
continuity of u across the element boundaries.
and quintics tend to oscillate unphysically, it is helpful to divide a large or complex

domain into smaller subdomains and use low-order piece-wise polynomials over each
of them – the subdomains again are the finite elements.
For example, a field variable u(x), may be represented by several linear elements as
illustrated in Fig. 4.1. It is generally necessary to impose constraints to ensure conti-
nuity of u across the element boundaries. Reparameterizing the linear function from
monomial coefficients a and b in one of the elements in terms of the nodal values of u
at each end of the element (u1 and u2 ), we write for one element
u(ξ ) = u1 (1 − ξ ) + u2 ξ, (4.38)
where ξ ∈ (0, 1) is a normalized measure of distance along the one-dimensional element
(Fig. 4.1).
Adjacent elements share global nodal parameters, U∆ (see Fig. 4.2), defined at each
global node, ∆. Thus, it is necessary to map global nodal parameters, U∆ , defined at
global node ∆, onto local node n of element e by use of a connectivity matrix, ∆(n, e),
un = U∆(n,e) . (4.39)
It is often desirable to use an interpolation that preserves continuity of the derivative of
a field variable u with respect to ξ across element boundaries not only u itself. Then be-
cause neighboring elements may not have the same size in the global coordinate system,
it is more accurate to define the global nodal derivative parameters as (∂u/∂s)n where
s is arc length in global units, and then to compute the local basis function parameters
(∂u/∂ξ )n , for each element with respect to its own local coordinates ξ ,

∂u ∂u ∂s
= · . (4.40)
∂ξ n ∂s ∆(n,e) ∂ξ n
F IG . 4.2. The finite element method: relationship between global nodal parameters Ui (i = 1, . . . , 4) and
local nodal parameters uj (j = 1, 2).
F IG . 4.3. Scaling factors.
The term (∂s/∂ξ )n is an element scale factor, which scales the arc length derivative of
global node ∆ to the local derivative of element node n required for the local interpola-
tion (Fig. 4.3).
4.4.2. Element parameterization

Schemes for discretizing a domain into a finite number of elements involve choosing
points within the domain that define each element and result in a mesh of nodes. (Vari-
ous approaches and rules for creating meshes can be found elsewhere.) Since each finite
element will be treated separately, similar calculations are repeated for each element. To
make the process more convenient each element is mapped to a reference element, a ba-
sic change of coordinates. For example, the geometry of an element that is shaped like
a thin plate of arbitrary size that may be warped out of plane with curved and skewed
boundaries in actual space can be mapped to a simple square with sides one unit long
(Fig. 4.4).
Similar to interpolation of a field variable, the geometry of Ω is then interpolated
over emax elements as
e
max M(e)

X= Ψ m Xm e, (4.41)
e=1 m=1
where M(e) = nmax (e) × γmax , the number of nodes defining the eth element times
the number of parameters per node. [Ψ m ] is matrix of interpolation functions defined
F IG . 4.4. Mapping of a geometric variable from the global to element coordinate system.
as products of independent Lagrange and Hermite polynomials in each ξk direction

for each node; e = (e1 , e2 , e3 ) are unit vectors, which describe a curvilinear coordinate
system; and [Xm ] is a matrix-vector of nodal coordinates and generalized derivatives.
Generalized derivatives are included in [Xm ], as in the case of the interpolation of a field
variable, for continuity of derivatives. Derivatives of the dependent variables may then
appear in an integrand, so they must also be transformed from the global to the local
coordinate system.
A scaling factor [Sm ] matrix is also defined for each element, so that equations may
be written
e
max M(e)

X= Ψ m Sm X∆(m,e) e, (4.42)
e=1 n=1
where [Xm ] = [Sm ][X∆(m,e) ].

Thus, both the geometric or dependent field variables of an element may in general
be interpolated in two or three dimensions by equations of the form
n
max γ max
γ γ
û(e) = Ψn(e) (ξ1 , ξ2 , ξ3 ) · un(e) , (4.43)
n(e)=1 γ =1
γ γ
where Ψn(e) (ξ1 , ξ2 , ξ3 ) are piece-wise interpolation functions, un(e) are nodal parameters
at local node n of element (e), and the index γ identifies each of the γmax nodal parame-
ters used to interpolate u(e) . For three-dimensional elements, nmax = 8. The interpola-
γ
tion functions Ψn(e) (ξ1 , ξ2 , ξ3 ) may be constructed as tensor products of separate poly-
nomials in each ξk direction. For three-dimensional rectangular finite elements, these
equations may be written as follows:
γ γ γ γ
Ψn(e) (ξ1 , ξ2 , ξ3 ) = ϕn(e)1 (ξ1 ) ϕn(e)2 (ξ2 ) ϕn(e)3 (ξ3 ). (4.44)
A full description of several Lagrange and Hermite interpolation functions and their
combinations in two and three dimensions can be found in Section 8 (Appendix).
5. Anatomical models
5.1. Coordinate systems
The geometry of a region, over which a finite element solution is sought, is defined
with respect to a coordinate system. The choice of coordinate system depends on the
particular problem. Fig. 5.1 shows the relation of Cartesian coordinates to coordinates
F IG . 5.1. The relationship between (A) rectangular Cartesian reference coordinates, Yi , and curvilinear
world coordinates, Θi , for two orthogonal coordinate systems that may be used to formulate the finite el-
ement equations: (B) cylindrical polar; and (C) spherical polar. Reprinted with permission of the ASME from
C OSTA , H UNTER , ROGERS , G UCCIONE , WALDMAN and M C C ULLOCH [1996] ASME J. of Biomech. Eng.,
118:452–463.
F IG . 5.2. The relationship between rectangular Cartesian reference coordinates, Yi , and prolate spheroidal
coordinates, {ΘA } = {Λ, M, Θ} used here to describe a thick-walled confocal ellipsoidal shell bounded by
inner and outer surfaces of constant Λ (the dimensionless transmural coordinate) and truncated at M = 120◦ .
Dimensional scaling is determined by the focal length, d. Reprinted with permission of the ASME from
C OSTA , H UNTER , WAYNE , WALDMAN , G UCCIONE and M C C ULLOCH [1996] ASME J. of Biomech. Eng.,
118:464–472.
F IG . 5.3. Four coordinate systems are used in our finite element method. A rectangular Cartesian global
reference coordinate system {Y1 , Y2 , Y3 } and orthogonal curvilinear coordinate systems {Θ1 , Θ2 , Θ3 } are
used to describe the geometry. Curvilinear local finite element coordinates are {ξ1 , ξ2 , ξ3 }, and locally ortho-
normal convecting body/fiber coordinates are {X1 , X2 , X3 }. Reprinted with permission of the ASME from
C OSTA , H UNTER , ROGERS , G UCCIONE , WALDMAN and M C C ULLOCH [1996] ASME J. of Biomech Eng.,
118:452–463.
F IG . 5.4. The fibrous structure of the anisotropic myocardium can be defined using locally orthonormal
coordinates defined by the tissue fiber-sheet microstructure and derived in terms of the local finite element
coordinates {ξA }.
of other common coordinate systems. It is sometimes more efficient and convenient

to use an orthogonal curvilinear coordinate system. The prolate spheroidal coordinate
system shown in Fig. 5.2 is useful for finite element modeling of the ventricles.
In the case of the ventricles, position vectors R = Y R eR are defined with respect to the
global rectangular Cartesian reference coordinate system, YA , or a suitable curvilinear
system of world coordinates, ΘA . The geometry of the body is discretized, and nodal
geometric variables are interpolated using polynomial functions of normalized finite
element coordinates ξk (Fig. 5.3).
The fibrous structure of the anisotropic myocardium is defined using locally ortho-
normal body/fiber coordinates Xi in which Xf is aligned with the local muscle fiber
axis and lies in the epicardial tangent coordinate plane (the ξ1 –ξ2 plane). Xs lies in the
laminar sheet coordinate plane, and Xn is orthogonal to the sheet plane (Fig. 5.4). The
relationship between these coordinate systems is described by a transformation matrix
[M], as shown by U SYK , M AZHARI and M C C ULLOCH [2000].
5.2. Least squares fitting for nodal geometric parameters
To apply the general theory of least squares fitting to fit finite element models to anatom-
ical measurements, we introduce the objective function

D
2
F (X) = γd X(ξd ) − Xd , (5.1)
d=1
where Xd is the dth measured coordinate or field variable, X(ξd ) is the interpolated
value at ξd , which is defined by the projection of the measured point onto a surface,
and γd is the corresponding weight applied to the data point. The objective function
represents the error between the coordinate of a measured anatomical surface point and
the corresponding coordinate projected on the element surface. Weight parameters are
all set equal to one when all measurements can be assumed to be equally accurate. As
shown above, we can define X as
e
max M(e)

X= Ψ m Sm X∆(m,e) e, (5.2)
e=1 m=1
so

D
emax M(e)
2
F (X) = γd Ψ m Sm X∆(m,e) e − Xd . (5.3)
d=1 e=1 m=1
A least squares fit minimizes the objective function
∂F
= 0. (5.4)
∂Xj∆(m,e)
The following sections describe various anatomical features that have been incorporated
into ventricular models using least squares fitting techniques.
5.3. Ventricular geometry
The mammalian heart consists of four pumping chambers, the left and right atria and
ventricles communicating through the atrioventricular (mitral and tricuspid) valves,
which are structurally connected by chordae tendineae to papillary muscles that ex-
tend from the anterior and posterior aspects of the right and left ventricular lumens. The
muscular cardiac wall is perfused via the coronary vessels that originate at the left and
right coronary ostia located in the sinuses of Valsalva immediately distal to the aortic
valve leaflets. Surrounding the whole heart is the collagenous parietal pericardium that
fuses with the diaphragm and great vessels.
From the perspective of engineering mechanics, the ventricles are three-dimensional
thick-walled pressure vessels with substantial variations in wall thickness and principal
curvatures both regionally and temporally through the cardiac cycle. The ventricular
walls in the normal heart are thickest at the equator and base of the left ventricle and
thinnest at the left ventricular apex and right ventricular free wall. There are also vari-
ations in the principal dimensions of the left ventricle with species, age, phase of the
cardiac cycle, and disease.
Ventricular geometry has been studied in most quantitative detail in the dog heart
(N IELSEN , L E G RICE , S MAILL and H UNTER [1991]). Geometric models have been
very useful in the analysis, especially the use of confocal and nonconfocal ellipses of
revolution to describe the epicardial and endocardial surfaces of the left and right ven-
tricular walls. The canine left ventricle is reasonably modeled by a thick ellipsoid of
revolution truncated at the base. The crescentic right ventricle wraps about 180 degrees
around the heart wall circumstantially and extends longitudinally about two thirds of the
distance from the base to the apex. Using a truncated ellipsoidal model, left ventricular
geometry in the dog can be defined by the major and minor radii of two surfaces, the
left ventricular endocardium, and a surface defining the free wall epicardium and the
septal endocardium of the right ventricle. S TREETER J R . and H ANNA [1973] described
the position of the basal plane using a truncation factor fb defined as the ratio between
the longitudinal distances from equator-to-base and equator-to-apex. Hence, the overall
longitudinal distance from base to apex is (1 + fb ) times the major radius of the ellipse.
Since variations in fb between diastole and systole are relatively small (0.45 to 0.51),
they suggested a constant value of 0.5.
The focal length d of an ellipsoid is defined from the major and minor radii (a and b)
by d 2 = a 2 − b 2 , and varies only slightly in the dog from endocardium to epicardium
between end-diastole (37.3 to 37.9 mm) and end-systole (37.7 to 37.1 mm) (S TREETER
J R . and H ANNA [1973]). Hence, within measurement accuracy, the boundaries of the
left ventricular wall can be treated as ellipsoids of revolution, and the assumption that
the ellipsoids are confocal appears to be a good one. This has motivated the choice
of prolate spheroidal (elliptic–hyperbolic–polar) coordinates (λ, µ, θ ) as described ear-
lier (N IELSEN , L E G RICE , S MAILL and H UNTER [1991], YOUNG and A XEL [1992]).
Here, the focal length d defines a family of coordinate systems that vary from spherical
polar when d = 0 to cylindrical polar in the limit when d → ∞. A surface of constant
transmural coordinate λ is an ellipse of revolution with major radius a = d cosh λ and
minor radius b = d sinh λ. In an ellipsoidal model with a truncation factor of 0.5, the
longitudinal coordinate µ varies from 0◦ at the apex to 120◦ at the base. Integrating the
Jacobian in prolate spheroidal coordinates gives the volume of the wall or cavity
2π µ λ2

d 3
(sinh2 λ + sin2 µ) sinh λ sin µ dλ dµ dθ
0 0 λ1
2πd 3

= (1 − cos µ) cosh3 λ − 1 − cos3 µ cosh λλ2 . (5.5)
3 λ
1
Using a truncated ellipsoidal model (Fig. 5.5) as an initial approximation and using
the finite element least squares fitting approach described above, V ETTER and M C -
C ULLOCH [1998] built a realistic anatomical model of the geometry of the right and
left ventricles of the rabbit heart (Fig. 5.6). By using prolate spheroidal coordinates to
F IG . 5.5. Initial unfitted prolated spheroidal meshes for the epicardial surface and the left and right ventric-
ular endocardium. The mesh represents the volume occupied by the left and right endocardial free walls and
the septal wall.
F IG . 5.6. Fitted model of the rabbit heart with epicardial and endocardial surfaces of the left and right ven-
tricles rendered (from V ETTER and M C C ULLOCH [1998]).
F IG . 5.7. Mesh fitted to porcine atrial anatomy.
construct surfaces as initial estimates for the left and right ventricular epicardia and
endocardia (Fig. 5.5), V ETTER and M C C ULLOCH [1998] reduced the problem to a
one-dimensional least squares fit of the λ coordinate alone, which was approximated on
each surface using bicubic Hermite interpolation after the original work of N IELSEN ,
L E G RICE , S MAILL and H UNTER [1991].
Using the same general techniques, it is possible to fit anatomical models to mea-
surements of other three-dimensional structures such as the atria and blood vessels, etc.
Fig. 5.7 shows an anatomical model of the porcine atria.
5.4. Fiber architecture
The cardiac ventricles have a complex three-dimensional muscle fiber architecture (for
a comprehensive review see S TREETER J R . [1979]). Although the myocytes are rela-
tively short, they are connected such that at any point in the normal heart wall there is
a clear predominant fiber axis that is approximately tangent with the wall (within 3–5◦
in most regions, except near the apex and papillary muscle insertions). Each ventricular
myocyte is connected via gap junctions at intercalated disks to an average of 11.3 neigh-
bors, 5.3 on the sides and 6.0 at the ends (S AFFITZ , K ANTER , G REEN , T OLLEY and
B EYER [1994]). The classical anatomists dissected discrete bundles of fibrous swirls,
though later investigations showed that the ventricular myocardium could be unwrapped
by blunt dissection into a single continuous muscle “band” (T ORRENT-G UASP [1973]).
However, more modern histological techniques showed that in the plane of the wall,
the muscle fiber angle makes a smooth transmural transition from epicardium to endo-
cardium. Similar patterns have been described for humans, dogs, baboons, macaques,
pigs, guinea pigs, and rats. In the human or dog left ventricle, the muscle fiber angle
typically varies continuously from about −60◦ (i.e., 60◦ clockwise from the circumfer-
ential axis) at the epicardium to about +70◦ at the endocardium. The rate of change of
fiber angle is usually greatest at the epicardium, so that circumferential (0◦ ) fibers are
found in the outer half of the wall, and the rate of angle change begins to slow trans-
murally approaching the inner third of the wall near the trabeculata–compacta interface.
There are also small increases in fiber orientation from end-diastole to systole (7–19◦),
with the greatest changes at the epicardium and apex (S TREETER J R ., S POTNITZ , PA -
TEL , ROSS J R . and S ONNENBLICK [1969]).
A detailed description of the morphogenesis of the muscle fiber system in the devel-
oping heart is not available, but there is evidence of an organized myofiber pattern by
day 12 in the fetal mouse heart that is similar to that seen at birth (day 20) (M C L EAN ,
ROSS and P ROTHERO [1989]). Abnormalities of cardiac muscle fiber patterns have been
described in some disease conditions. In hypertrophic cardiomyopathy, which is often
familial, there is substantial myofiber disarray, typically in the interventricular septum
(M ARON , B ONOW, C ANNON 3 RD , L EON and E PSTEIN [1987]).
Regional variations in ventricular myofiber orientations are generally smooth except
at the junction between the right ventricular free wall and septum. A detailed study in
the dog that mapped fiber angles throughout the entire right and left ventricles described
the same general transmural pattern in all regions including the septum and right ven-
tricular free wall, but with definite regional variations (N IELSEN , L E G RICE , S MAILL
and H UNTER [1991]). Transmural differences in fiber angle were about 120–140◦ in the
left ventricular free wall, larger in the septum (160–180◦), and smaller in the right ven-
tricular free wall (100–120◦). A similar study of fiber angle distributions in the rabbit
left and right ventricles has recently been reported (V ETTER and M C C ULLOCH [1998]).
Fiber angles in the rabbit heart were generally very similar to those in the dog, except
for on the anterior wall, where average fiber orientations in the rabbit were 20–30◦
counterclockwise of those in the dog.
F IG . 5.8. Anatomical model of the rabbit left and right ventricles. 8,351 geometric points and 14,368 fiber
angles were fitted using 36 high-order finite element elements.
F IG . 5.9. Fitted fiber angles in the lateral left ventricular wall: + experimental measurements (rabbit); — fit-
ted data (rabbit) (V ETTER and M C C ULLOCH [1998]); – – fitted data (dog) (N IELSEN , L E G RICE , S MAILL
and H UNTER [1991]).
Using the same least squares method that was used to fit the ventricular geometry de-
scribed in the previous section, V ETTER and M C C ULLOCH [1998] also fitted a model
of fiber architecture into the anatomical model of the rabbit heart. The fitted model
was based on about 14,000 histologically measured angles (Fig. 5.8). Figs. 5.9 and
5.10 show experimental measurements and model values for fiber angles in different
F IG . 5.10. Fitted fiber angles in the anterior wall: + local experimental measurements (rabbit); — fitted
data (rabbit) (V ETTER and M C C ULLOCH [1998]); – – fitted data (dog) (N IELSEN , L E G RICE , S MAILL and
H UNTER [1991]).
regions of the left ventricular wall from a dog model (N IELSEN , L E G RICE , S MAILL
and H UNTER [1991]) and a rabbit model (V ETTER and M C C ULLOCH [1998]). Typ-
ical root-mean-squared fitting errors were less than 15–20◦ which is in the range of
measurement error.
5.5. Sheet architecture
The fibrous architecture of the myocardium has motivated models of myocardial mate-
rial symmetry as transversely isotropic. The recognition by L E G RICE , S MAILL , C HAI ,
E DGAR , G AVIN and H UNTER [1995] that planes of cleavage observed in transverse
myocardial sections correspond to parallel, branching laminar sheets several myocytes
thick are the best structural evidence for material orthotropy and have motivated the
development of models describing the variation of fiber, sheet, and sheet-normal axes
throughout the ventricular wall (L E G RICE , H UNTER and S MAILL [1997]). This also
led to the hypothesis that the laminar architecture of ventricular myocardium is related
to significant transverse shear strains (WALDMAN , F UNG and C OVELL [1985]) and my-
ofiber rearrangement (S POTNITZ , S POTNITZ , C OTTRELL , S PIRO and S ONNENBLICK
[1974]) observed in the intact heart during systole. By measuring three-dimensional dis-
tributions of strain across the wall thickness using biplane radiography of radiopaque
markers, LeGrice and colleagues (L E G RICE , TAKAYAMA and C OVELL [1995]) found
that the cleavage planes coincide closely with the planes of maximum shearing during
ejection, and that the consequent reorientation of the myocytes may contribute 50% or
more of normal systolic wall thickening. For a discussion of the implications of sheet
organization on electrical propagation see Section 3.1.
5.6. Conductivity
Conductivity in the intracellular and extracellular spaces is represented by the tensors

Di and De . The use of tensors allows directions of preferential conductivity to be de-
fined. Since the tissue is known to have higher conductivity in the direction of the fibers
than in other directions and is hypothesized to also have higher conduction in the di-
rection of sheets than in the sheet-normal direction, the conductivity tensor is defined
relative to the fiber–sheet coordinate system (see Fig. 5.4). When in this coordinate sys-
tem, the conductivity tensor is diagonalized. Its highest eigenvalue corresponds to the
eigenvector in the direction of the fiber coordinate direction. Because fiber and sheet
angles vary regionally in the tissue, the directions of fastest conductivity will also vary
in the same way. Conductivity can then be defined as a constant ratio of fiber to sheet to
sheet normal magnitudes with only the orientation of the tensors varying regionally.
Recently least squares interpolation methods have also been used to incorporate a
Purkinje fiber network into an electrophysiological model of the dog ventricles (U SYK ,
L E G RICE and M C C ULLOCH [2002]). Measurements of the geometry of the Purkinje
fibers on the endocardium were used to fit a field representing the local angle of the
Purkinje network. This field was defined for the whole endocardium even though Purk-
inje fibers are found only in discrete locations. A second field was fitted to represent
the faster conduction that occurs along the Purkinje fibers. This field is defined with the
constraint that its components are non-zero only in locations corresponding to positions
where a fiber can be found. These two fields together define an additional conductivity
tensor that can be included by superposition.
6. Solution implementation
The bidomain equations with the associated natural boundary conditions are summa-
rized as
dVm
∇ · Di ∇Vm + ∇ · Di ∇ue = Sv Cm + Sv Iion in Ω, (6.1)
dt
∇ · (Di + De )∇ue = −∇ · Di ∇Vm in Ω, (6.2)
n · Di ∇Vm = 0 on ΓE . (6.3)
Eqs. (6.1) and (6.2) must be solved simultaneously, so that one of each of the two un-
knowns, Vm and ue , can be found from each of the equations. Collecting terms involving
Vm on the left-hand side of the first equation:
dVm
Sv Cm − ∇ · Di ∇Vm = ∇ · Di ∇ue − Sv Iion . (6.4)
dt
Applying weighted residuals as in Eq. (4.28):

dVm
Ψk Sv Cm dΩ − ∇Ψk · Di ∇Vm dΩ
Ω dt Ω

= Ψk (∇ · Di ∇ue − Sv Iion ) dΩ, k = 1, . . . , M, (6.5)
Ω
where the boundary conditions have been used to cancel out the surface integral term
from the integration by parts of the second term on the left-hand side. Using integration
by parts again on the first term on the right-hand side prevents any second derivative
terms from entering the equation:
M

d(Ψj Vmj )
Ψk Sv Cm dΩ − ∇Ψk · Di ∇Ψj dΩ Vmj
Ω dt Ω
j =1

= Ψk Di ∇ue · n dΓ − ∇Ψk · Di ∇ue dΩ − Ψk Sv Iion dΩ. (6.6)
Γ Ω Ω
Eq. (6.2) is already written with terms involving ue on the left-hand side. Similarly
applying weighted residuals with natural boundary conditions:
M

∇Ψk · (Di + De )∇Ψj dΩ uej = − Ψk (∇ · Di ∇Vm ) dΩ. (6.7)
j =1 Ω Ω
In both equations k = 1, . . . , M.
Since the integral over the whole domain is the sum of integrals over each subdo-
main, the finite element method provides a mechanism not only for solving the numer-
ical problem but for assembling it. Discretizing the domain into emax elements, a pair
of equations for each element subdomain can be written whose sum is the original
Eqs. (6.6) and (6.7):
M

d(Ψj Vmj )
Ψk Sv Cm dΩ −
e
∇Ψk · Di ∇Ψj dΩ Vmj
e
Ωe dt Ωe
j =1

= Ψk Di ∇ue · n dΓ e − ∇Ψk · Di ∇ue dΩ e − Ψk Sv Iion dΩ e , (6.8)
Γe Ωe Ωe
M

∇Ψk · (Di + De )∇Ψj dΩ e uej = − Ψk (∇ · Di ∇Vm ) dΩ e , (6.9)
j =1 Ωe Ωe
where an M × M system of equations can be constructed separately for each element.

Because elements share global nodes, the system of equations for each element cannot
be solved independently. Hence they are assembled into a single whole system for all
the elements where there is one row for each global nodal parameter. The connectivity
matrix dictates which local nodal parameters (rows in element M × M systems) corre-
spond to the same global parameter and should thus be summed on a single row of the
whole system.
At this stage, the anatomy of the ventricles is incorporated through the representation
of the geometry of the elements in the domain. The structural properties of tissue archi-
tecture are incorporated through the combination of fields representing fiber geometry,
sheet geometry, and conductivity in the terms Di and De . Local cellular characteristics
are incorporated through the membrane currents calculated in the term Iion . The next
section describes methods that have been used to solve these equations.
6.1. Solution algorithms
The bidomain equations must be discretized in time as well as space. Finite difference
schemes based on the θ -rule are commonly used. The θ -rule for a problem that is first
order in time
du
= G, (6.10)
dt
where G is some spatially varying function of u, is written as
un+1 − un
= θ Gn+1 + (1 − θ )Gn , (6.11)
t
where values of un are known, un+1 is the unknown quantity, and time has been dis-
cretized into steps of size t. Choices of θ lead to various methods, for example,
θ = 0 → forward Euler, θ = 12 → Crank–Nicolson, θ = 1 → backward Euler.
Various schemes have been used for discretizing the bidomain problems in time. For
example, S UNDNES , L INES and T VEITO [2001] have recently used operator splitting
methods to solve the bidomain equations simultaneously with PDEs representing the
potential in the torso. Their calculations are useful for investigating the forward prob-
lem of electrophysiology, which refers to the calculation of body surface potentials
from potentials originating in the heart as are measured clinically through electrocar-
diograms. Operator splitting methods consist of separating the ODE calculations from
the PDE calculations by dividing a single time step. During the first part of the time step
the ODEs are solved. Their solution is used to update the PDEs before solving them
during the second half time step. In this way the Iion term becomes a constant source
term rather than a function of Vm during the solution of the PDEs with a value from the
solution of the ODEs at the half time step.
In the discretized bidomain equations, Eq. (6.8) is first order in time due to the
dVm /dt term on the left-hand side. This term can be discretized as follows:

d(Ψj Vmj ) Sv Cm
n+1
Sv Cm Ψk dΩ =
e
Ψk Ψj Vmj − Vmn dΩ e , (6.12)
Ω e dt Ω e t
where Vmn is known and Vmn+1 is the variable for which an approximation is being
sought. Employing the operator splitting method of Sundnes et al.: treating the Iion
term as known from a series of separate integrations of the ODEs over the global time
interval from t = n to t = n + 1/2, Eqs. (6.8) and (6.9) can be written as
M

Sv Cm n+1
Ψk Ψj − ∇Ψk · Di ∇Ψj Vmj dΩ e
Ωe t
j =1

Sv Cm n+1/2
= Ψk Ψj Vm − ∇Ψk · Di ∇ue − Ψk Sv Iion
n n
dΩ e
Ωe t

+ Ψk Di ∇une · n dΓ e , (6.13)
Γe
M

∇Ψk · (Di + De )∇Ψj dΩ e un+1
ej = − Ψk ∇ · Di ∇Vmn+1 dΩ e . (6.14)
j =1 Ωe Ωe
The steps for finding the solution proceed as follows. First, the Iion term is found
by separate integration of the ODEs at each integration point in the discretized mesh
representing the ventricles using known quantities at time n as the initial conditions.
Next, the results of the ODE integration at time n + 1/2 are combined with the known
solutions of Vm and ue at time n to find the terms in the right-hand side of the first
bidomain equation. Following assembly, a solution of a linear system of the form
Ax = b, (6.15)
is solved for Vmn+1 at each of the M global nodes. Finally, Vmn+1 is used to form the right-
hand side of the second equation and a second assembly and linear system solution is
performed to find un+1
e at the M global nodes.
6.2. Implementation issues
Although computer memory and speed continue to grow with advances in technology,
simulations of cardiac electrophysiology problems remain large and time consuming
and continue to rapidly grow more so with advances in biology. Parallel programming
methods are useful in this situation. In incorporating local cellular processes into the
simulation of tissue electrophysiology, a level of data parallelism can be achieved. If
local values of membrane potential are known, the currents passing through the ion
channels or within the intracellular spaces of a cell in one location can be calculated
independently of those occurring within a cell in another location. In time integration
schemes such as the method presented in the previous section, this is the exact situation.
The result is that cellular information is data parallel. Calculations of local cellular
processes can be distributed to a set of processors so that multiple calculations can be
performed simultaneously. As the ODE calculation time grows with the sophistication
of the ionic model, the solution of the linear systems do not change size or complexity.
Then for complex ionic models, the data parallel ODE portion of the solution is also the
portion where improvements in speed are most needed.
The nature of the propagation problem is that local areas of tissue near the wave front
are the same areas experiencing fast ion kinetics. In locations far from the wave front,
kinetics are relatively slow. Parallel programming that balances the calculation load
could lead to even further speedups. Load balancing works by keeping processors busy,
for example, while one processor calculates an ODE integration for a point near the
wave front requiring many small time steps to cover the global time increment, another
processor might calculate ODE integrations for several points away from the wave front
that each require few local time steps to cover the same global time increment.
Adaptive meshing techniques involving element size can also be used for improving
efficiency. These techniques use smaller mesh elements to discretize the domain in re-
gions near the wave front and larger elements in areas away from the wave front. For
adaptive meshing and load balancing techniques the location of the wave front must be
identified. In addition, for adaptive meshing the wave front location must also be pre-
dicted so that it will not reach elements that are too large in a single time step leading to a
divergent solution. Various adaptive meshing techniques applied to cardiac electrophys-
iology can be found in Q U and G ARFINKEL [1999], Q UAN , E VANS and H ASTINGS
[1998], OTANI [2000] and elsewhere.
In examining the solution of cardiac propagation problems various major problem
components can be identified each of which may be accomplished by significant sep-
arate software developments. For example, the ODE integration of the cellular ionic
model may be carried out by an implicit Runge–Kutta solver suitable for stiff problems,
but there are many equally effective choices of solvers. The best one to use may depend
on exactly which ionic model has been chosen. Interchangeability of the ionic model
itself may also be important since these models are evolving rapidly. Similarly, the so-
lution of linear systems for the PDEs may be performed by a range of Ax = b solvers.
The best solver may depend on the number of mesh elements or the computational plat-
form being utilized, i.e., some solvers use matrix free methods that can solve the linear
system without assembling A resulting in less memory use while other solvers may be
specialized for solving linear systems on a distributed memory multiprocessor system.
Several relatively new paradigms in computing can address these implementation re-
quirements. Object-oriented programming treats software components as objects, black
boxes that send and receive messages. For example, an ODE integrator object would be
sent a message describing which equations to integrate, when to start, and when to stop
integrating. All ODE objects need this same type of information, and all of them re-
sult in values of the dependent variables of the ODEs calculated over time. Every time a
user has a new set of ODEs to integrate, it is not necessary to change the ODE integrator
object itself in order to integrate the new set of equations.
The use of generic programming methods in software developments can help facili-
tate the adaptability and interoperability of separately developed software objects. The
goal of generic programming is to express components of programs at the most general
level possible without losing efficiency. This involves writing algorithms with mini-
mum assumptions about the data to be processed, while also creating data structures
with minimal assumptions about the algorithms to be used. This provides for maximum
interoperability of separate components. With generic programming, components can
be developed independently and combined arbitrarily as needed only requiring speci-
fied interfaces in order to communicate. Then treating the components as objects and
gluing them together in a component-based environment, should lead to faster software
development times where reuse of components and overall code modifiability can be
maximized. In addition, generic programming helps to reduce the number of lines of
code, thus reducing the possibility of bugs as well as lowering maintenance costs.
For the bidomain problem of electrophysiology various objects need to be used
together in order to solve the problem. Component-based software design can be
used for gluing objects together. Programming languages such as Python, a very high
level object-oriented programming language, can be used for this implementation task.
Python can interface code written in other languages and handle the conversion of data
structures so that various independent pieces of code can function together as objects in
an overall software package.
7. Integrated models
The electrophysiological function of the heart does not exist separately from its mechan-
ical function. So far the integration across biophysical scales has been discussed in the
context of incorporation of cellular ionic systems models into finite element simulations
of wave propagation in tissues. However, the incorporation of mechanical function into
such simulations is of fundamental interest. Contraction itself occurs within sarcom-
eres at the cellular level and depends on the intracellular calcium concentration (see
Section 2.4).
Three-dimensional finite element stress analysis methods for large elastic deforma-
tions of nonlinear anisotropic materials can be applied to modeling the mechanics of the
heart. For example, the steps needed to solve a mechanics problem may proceed as fol-
lows. Finite element equations are integrated using a Gaussian quadrature scheme, and
the resulting system of nonlinear elliptic equations are solved for the unknown deforma-
tion and pressure using a Newton iterative method (O DEN [1972]). The non-symmetric
element tangent stiffness matrix (Jacobian) may be approximated by forward differ-
ences or may be found analytically and updated at each full Newton iteration. Non-zero
contributions to the constraint-reduced global tangent stiffness matrix are vectorized
and solved using a general linear sparse solver with threshold pivoting. The iterative
process is terminated when the sum of solution increments and the maximum uncon-
strained residual are both less than an acceptable threshold (U SYK , L E G RICE and M C -
C ULLOCH [2002]).
If the finite deformation stress analysis above incorporates cellular level informa-
tion in a manner similar to the electrophysiology problem, it is conceivable that bi-
directional influences of each of these large problems could be implemented through
local interactions of cellular models. The applications of this type of integration are clear
in measurable phenomena such as excitation–contraction coupling and mechanoelectric
feedback. Other systems models would also be useful extensions to the electrophysio-
logical problem. For example, signaling, metabolism, and energetics all represent fields
where important components could be collected for integration into a single model.
With these features, simulations of heart failure, ischemia, and other conditions, not to
mention normal physiological function will be more complete.
The computational hurdles for creating an integrated model include schemes for syn-
chronizing calculations and translating meshes since these problems are typically solved
with very different time and space scales. Parallelism can be exploited in an additional
layer in an integrated model where, for example, the electrical and mechanical portions
of the problem are solved on different processors with carefully developed commu-
nication between these major problem objects. Issues of convergence and parameter
sensitivity are also of increasing concern as the number of variables increases.
In summary, integration across biological scales in simulating cardiac electrophysi-
ology is common in current cardiac models. This paves the way for integration of struc-
turally and functionally integrated models of cardiac electromechanical function that
combine data-intensive cellular systems models with compute-intensive anatomically
detailed multiscale simulations.
8. Appendix: Lagrange and Hermite interpolations
8.1. Linear interpolation
For any variable u, a linear variation between two values, u1 and u2 , may be described
as
u(ξ ) = (1 − ξ )u1 + ξ u2 , 0 ξ 1, (8.1)

where the parameter ξ is a normalized measure of distance along the curve. (Notice
that in previous sections u represented the solution over an entire domain, and u(e)
represented the piece-wise solution on element subdomains:
e
max
u= u(e) . (8.2)
1
In this appendix, the (e) notation has been dropped for convenience and clarity so that
u is u(e) used in other sections.) We define
ϕ1 (ξ ) = 1 − ξ, ϕ2 (ξ ) = ξ, (8.3)
so that
u(ξ ) = ϕ1 (ξ )u1 + ϕ2 (ξ )u2 , (8.4)

where ϕ1 (ξ ) and ϕ2 (ξ ) are the linear Lagrange basis functions associated with the nodal
parameters u1 and u2 (Fig. 8.1).
8.2. Quadratic interpolation
A quadratic variation of u over a one-dimensional element requires three nodal parame-

ters,
u(ξ ) = ϕ1 (ξ )u1 + ϕ2 (ξ )u2 + ϕ3 (ξ )u3 . (8.5)

The quadratic Lagrange basis functions are shown in Fig. 8.2. Notice that since ϕ1 (ξ )
must be zero at ξ = 0.5 (node 2), ϕ1 (ξ ) must have a factor (ξ − 0.5) and since it is
also zero at ξ = 1 (node 3), another factor is (ξ − 1). Finally, since ϕ1 (ξ ) is 1 at ξ = 0
F IG . 8.1. The one-dimensional linear Lagrange basis functions.

F IG . 8.2. The quadratic Lagrange basis functions.
(node 1), we have
ϕ1 (ξ ) = 2(ξ − 0.5)(ξ − 1). (8.6)

The other two quadratic Lagrange basis functions are found similarly, so the quadratic
Lagrange basis functions are
ϕ1 (ξ ) = 2(ξ − 0.5)(ξ − 1),

ϕ2 (ξ ) = −4ξ(ξ − 1),
ϕ3 (ξ ) = 2ξ(ξ − 0.5). (8.7)
8.3. Cubic Hermite interpolation
All the basis functions mentioned thus far are Lagrange basis functions and provide
C 0 continuity of u across element boundaries but not higher-order continuity. In order
to preserve continuity of the derivative of u with respect to ξ across element boundaries
additional nodal parameters are included: the derivatives at node n, ( ∂u ∂ξ )n . The basis
functions are chosen to ensure that

∂u ∂u ∂u ∂u
= = u 1 and = = u2 , (8.8)
∂ξ ξ =0 ∂ξ 1 ∂ξ ξ =1 ∂ξ 2
and since u is shared between adjacent elements, derivative continuity is ensured. The
cubic Hermite basis functions are derived from
u(ξ ) = a + bξ + cξ 2 + dξ 3 , (8.9)
∂u
= b + 2cξ + 3dξ 2 , (8.10)
∂ξ
F IG . 8.3. Cubic Hermite basis functions.
subject to the constraints
u(0) = a = u1 ,
u(1) = a + b + c + d = u2 ,
∂u ∂u
(0) = b = u1 , (1) = b + 2c + 3d = u2 . (8.11)
∂ξ ∂ξ
Solving these equations, we get
u(ξ ) = ϕ11 (ξ )u1 + ϕ12 (ξ )u1 + ϕ21 (ξ )u2 + ϕ22 (ξ )u2 , (8.12)

where the four cubic Hermite basis functions are sketched in Fig. 8.3, and are given by
the following equations:
ϕ11 (ξ ) = 1 − 3ξ 2 + 2ξ 3 , ϕ12 (ξ ) = ξ(ξ − 1)2 ,

ϕ21 (ξ ) = ξ 2 (3 − 2ξ ), ϕ22 (ξ ) = ξ 2 (ξ − 1). (8.13)
8.4. Two-dimensional elements
Two-dimensional bilinear basis functions are readily constructed from the products of
the above one-dimensional linear functions as follows:

4
u(ξ1 , ξ2 ) = Ψn (ξ1 , ξ2 ) · un , (8.14)
n=1
where n are the four local nodes of a two-dimensional rectangular element (again the
notation indicating that these terms belong to a single element has been dropped, so
F IG . 8.4. Two-dimensional isoparametric element with linear Lagrange interpolation in one direction and
quadratic Lagrange interpolation in the other.
n = n(e) and u = u(e) of previous sections):

Ψ1 (ξ1 , ξ2 ) = ϕ1 (ξ1 )ϕ1 (ξ2 ), Ψ3 (ξ1 , ξ2 ) = ϕ1 (ξ1 )ϕ2 (ξ2 ),
Ψ2 (ξ1 , ξ2 ) = ϕ2 (ξ1 )ϕ1 (ξ2 ), Ψ4 (ξ1 , ξ2 ) = ϕ2 (ξ1 )ϕ2 (ξ2 ), (8.15)
and the functions ϕi (ξk ) (i, k = 1, 2) are defined above by Eqs. (8.3).
Higher-order, two-dimensional parametric basis functions can be similarly con-
structed from products of the appropriate one-dimensional basis functions. For exam-
ple, a six-noded (see Fig. 8.4) quadratic-linear element (quadratic in ξ1 and linear in ξ2 )
would be given by

6
u(ξ1 , ξ2 ) = Ψn (ξ1 , ξ2 ) · un , (8.16)
n=1
Q Q
Ψ1 (ξ1 , ξ2 ) = ϕ1 (ξ1 )ϕ1L (ξ2 ), Ψ4 (ξ1 , ξ2 ) = ϕ1 (ξ1 )ϕ2L (ξ2 ),
Ψ2 (ξ1 , ξ2 ) = ϕ2Q (ξ1 )ϕ1L (ξ2 ), Ψ5 (ξ1 , ξ2 ) = ϕ2Q (ξ1 )ϕ2L (ξ2 ),
Ψ3 (ξ1 , ξ2 ) = ϕ3Q (ξ1 )ϕ1L (ξ2 ), Ψ6 (ξ1 , ξ2 ) = ϕ3Q (ξ1 )ϕ2L (ξ2 ), (8.17)
where Q indicates quadratic basis functions as in Eqs. (8.7), and L indicates linear basis
functions as in Eqs. (8.3).
A two-dimensional bicubic Hermite element requires four derivatives per node for a
total of 16 parameters,
∂u ∂u ∂ 2u
u, , , and . (8.18)
∂ξ1 ∂ξ2 ∂ξ1 ∂ξ2
The need for the second order cross derivative term can be explained as follows: if u is
∂u ∂u
cubic in ξ1 and cubic in ξ2 then ∂ξ 1
is quadratic in ξ1 and cubic in ξ2 , and ∂ξ 2
is cubic
in ξ1 and quadratic in ξ2 . Now consider the 2–3 edge as shown in Fig. 8.5. The
cubic
∂u
variation of u with ξ2 is specified by four nodal parameters: two at node 1, u1 , ∂ξ

∂u ∂u
2 1
and two at node 3, u3 and ∂ξ 2 3
. Since ∂ξ1 (the normal derivative) is also cubic in ξ2
and is entirely independent of the four nodal parameters, four additional parameters are
∂u .
F IG . 8.5. Bicubic interpolation. Arrows on the 2–3 edge depict the direction of the normal derivatives, ∂ξ
1
required to specify that

∂u cubic
shape.
Two of these are specified by the normal derivatives
∂u
at nodes 1 and 3, ∂ξ and , and the remaining two by the variation of the normal
1 1 ∂ξ 1 3
2u
2u
derivatives with respect to ξ2 at nodes 1 and 3, ∂ξ∂1 ∂ξ 2 1
and ∂ξ∂1 ∂ξ 2 3
. The bicubic
interpolation is thus given by four nodal parameters at each node,

4
4
γ γ
u(ξ1 , ξ2 ) = Ψn (ξ1 , ξ2 ) · un , (8.19)
n=1 γ =1
where each node, n, of the element has the following nodal parameters:

∂u
u1n = un , u3n = ,
∂ξ2 n
2
∂u ∂ u
u2n = , u4n = , (8.20)
∂ξ1 n ∂ξ1 ∂ξ2 n
Ψ11 (ξ1 , ξ2 ) = ϕ11 (ξ1 )ϕ11 (ξ2 ), Ψ21 (ξ1 , ξ2 ) = ϕ21 (ξ1 )ϕ11 (ξ2 ),
Ψ31 (ξ1 , ξ2 ) = ϕ11 (ξ1 )ϕ21 (ξ2 ), Ψ41 (ξ1 , ξ2 ) = ϕ21 (ξ1 )ϕ21 (ξ2 ),
Ψ12 (ξ1 , ξ2 ) = ϕ12 (ξ1 )ϕ11 (ξ2 ), Ψ22 (ξ1 , ξ2 ) = ϕ22 (ξ1 )ϕ11 (ξ2 ),
Ψ32 (ξ1 , ξ2 ) = ϕ12 (ξ1 )ϕ21 (ξ2 ), Ψ42 (ξ1 , ξ2 ) = ϕ22 (ξ1 )ϕ21 (ξ2 ),
Ψ13 (ξ1 , ξ2 ) = ϕ11 (ξ1 )ϕ12 (ξ2 ), Ψ23 (ξ1 , ξ2 ) = ϕ21 (ξ1 )ϕ12 (ξ2 ),
Ψ33 (ξ1 , ξ2 ) = ϕ11 (ξ1 )ϕ22 (ξ2 ), Ψ43 (ξ1 , ξ2 ) = ϕ21 (ξ1 )ϕ22 (ξ2 ),
Ψ14 (ξ1 , ξ2 ) = ϕ12 (ξ1 )ϕ12 (ξ2 ), Ψ24 (ξ1 , ξ2 ) = ϕ22 (ξ1 )ϕ12 (ξ2 ),
Ψ34 (ξ1 , ξ2 ) = ϕ12 (ξ1 )ϕ22 (ξ2 ), Ψ44 (ξ1 , ξ2 ) = ϕ22 (ξ1 )ϕ22 (ξ2 ), (8.21)
j
and the functions ϕi (ξk ) (i = 1, 2; j = 1, 2; k = 1, 2, 3) are defined above by
Eqs. (8.13).
8.5. Three-dimensional elements
Three-dimensional trilinear Lagrange basis functions are similarly constructed from the
products of the above one-dimensional linear functions as follows:

8
u(ξ1 , ξ2 , ξ3 ) = Ψn (ξ1 , ξ2 , ξ3 ) · un , (8.22)
n=1
where
Ψ1 (ξ1 , ξ2 , ξ3 ) = ϕ1 (ξ1 )ϕ1 (ξ2 )ϕ1 (ξ3 ), Ψ5 (ξ1 , ξ2 , ξ3 ) = ϕ1 (ξ1 )ϕ1 (ξ2 )ϕ2 (ξ3 ),
Ψ4 (ξ1 , ξ2 , ξ3 ) = ϕ2 (ξ1 )ϕ2 (ξ2 )ϕ1 (ξ3 ), Ψ8 (ξ1 , ξ2 , ξ3 ) = ϕ2 (ξ1 )ϕ2 (ξ2 )ϕ2 (ξ3 )
(8.22a)
and the functions ϕi (ξk ) (i = 1, 2; k = 1, 2, 3) are defined above by Eqs. (8.3). These
eight basis functions correspond to the eight nodes of a trilinear brick element (Fig. 8.6).
A three-dimensional tricubic Hermite element requires eight derivatives per node,
∂u ∂u ∂u ∂ 2u ∂ 2u
u, , , , , ,
∂ξ1 ∂ξ2 ∂ξ3 ∂ξ1 ∂ξ2 ∂ξ1 ∂ξ3
∂ 2u ∂ 3u
, and , (8.23)
∂ξ2 ∂ξ3 ∂ξ1 ∂ξ2 ∂ξ3
8 8
γ γ
u(ξ1 , ξ2 , ξ3 ) = Ψn (ξ1 , ξ2 , ξ3 ) · un , (8.24)
n=1 γ =1
where

∂u ∂u
u1n = un , u2n = , u3n = ,
∂ξ1 n ∂ξ2 n

∂ 2u ∂u ∂ 2u
u4n = , u5n = , u6n = ,
∂ξ1 ∂ξ2 n ∂ξ3 n ∂ξ1 ∂ξ3 n
F IG . 8.6. An 8-noded three-dimensional isoparametric finite element.


∂ 2u ∂ 3u
u7n = , u8n = , (8.25)
∂ξ2 ∂ξ3 n ∂ξ1 ∂ξ2 ∂ξ3 n
Ψ78 (ξ1 , ξ2 , ξ3 ) = ϕ12 (ξ1 )ϕ22 (ξ2 )ϕ22 (ξ3 ), Ψ88 (ξ1 , ξ2 , ξ3 ) = ϕ22 (ξ1 )ϕ22 (ξ2 )ϕ22 (ξ3 ).
j
The functions ϕi (ξk ) (i = 1, 2; j = 1, 2; k = 1, 2, 3) are defined above by Eqs. (8.13).
References
A KAR , F.G., YAN , G.X., A NTZELEVITCH , C., ROSENBAUM , D.S. (2002). Unique topographical distrib-
ution of M cells underlies reentrant mechanism of torsade de pointes in the long-QT syndrome. Circula-
tion 105 (10), 1247–1253.
A LBERTS , B., B RAY , D., L EWIS , J., R AFF , M., ROBERTS , K., WATSON , J. (1994). Molecular Biology of
the Cell, third ed. (Garland Publishing, Inc., New York).
A NSARI , A., H O , S.Y., A NDERSON , R.H. (1999). Distribution of the Purkinje fibres in the sheep heart. Anat.
Rec. 254 (1), 92–97.
A NTZELEVITCH , C., S HIMIZU , W., YAN , G.X., S ICOURI , S., W EISSENBURGER , J., N ESTERENKO , V.V.,
B URASHNIKOV , A., D I D IEGO , J., S AFFITZ , J., T HOMAS , G.P. (1999). The M cell: its contribution to
the ECG and to normal and abnormal electrical function of the heart. J. Cardiovasc. Electrophysiol. 10 (8),
1124–1152.
A NYUKHOVSKY , E.P., S OSUNOV, E.A., ROSEN , M.R. (1996). Regional differences in electrophysiological
properties of epicardium, midmyocardium, and endocardium. In vitro and in vivo correlations. Circula-
tion 94 (8), 1981–1988.
A RTS , T., R ENEMAN , R.S., V EENSTRA , P.C. (1979). A model of the mechanics of the left ventricle. Ann.
Biomed. Eng. 7, 299–318.
A RTS , T., V EENSTRA , P.C., R ENEMAN , R.S. (1982). Epicardial deformation and left ventricular wall me-
chanics during ejection in the dog. Am. J. Physiol. Heart Circ. Physiol. 243, H379–H390.
BACKX , P.H., G AO , W.D., A ZAN -BACKX , M.D., M ARBAN , E. (1995). The relationship between contrac-
tile force and intracellular [Ca2+] in intact rat cardiac trabeculae. J. Gen. Physiol. 105 (1), 1–19.
B EELER , G.W., R EUTER , H. (1977). Reconstruction of the action potential of ventricular myocardial fibres.
J. Physiol. 268 (1), 177–210.
B ERNE , R.M., L EVY , M.N. (1997). Cardiovascular Physiology, seventh ed. (Mosby–Year Book, Inc.,
St Louis).
B ERNUS , O., W ILDERS , R., Z EMLIN , C.W., V ERSCHELDE , H., PANFILOV, A.V. (2002). A computation-
ally efficient electrophysiological model of human ventricular cells. Am. J. Physiol. Heart Circ. Phys-
iol. 282 (6), H2296–H2308.
B ERS , D.M. (1991). Excitation–contraction Coupling and Cardiac Contractile Force (Kluwer, Dordrecht).
C AREY , G.F., O DEN , J.T. (1983). Finite Elements, A Second Course, vol. II (Prentice Hall, Englewood
Cliffs, NJ).
C ATES , A.W., S MITH , W.M., I DEKER , R.E., P OLLARD , A.E. (2001). Purkinje and ventricular contribu-
tions to endocardial activation sequence in perfused rabbit right ventricle. Am. J. Physiol. Heart Circ.
Physiol. 281 (2), H490–H505.
C HADWICK , R.S. (1982). Mechanics of the left ventricle. Biophys. J. 39 (3), 279–288.
C HENG , J., K AMIYA , K., L IU , W., T SUJI , Y., T OYAMA , J., K ODAMA , I. (1999). Heterogeneous distribution
of the two components of delayed rectifier K + current: a potential mechanism of the proarrhythmic effects
of methanesulfonanilideclass III agents. Cardiovasc. Res. 43 (1), 135–147.
C IARLET , P.G., L IONS , J.L. (eds.) (1990). Finite Element Methods (Part 1). In: Handbook of Numerical
Analysis II (North-Holland, Amsterdam).
C OSTA , K.D., H UNTER , P.J., ROGERS , J.M., G UCCIONE , J.M., WALDMAN , L.K., M C C ULLOCH , A.D.
(1996). A three-dimensional finite element method for large elastic deformations of ventricular my-
ocardium: Part I – Cylindrical and Spherical Polar Coordinates. ASME J. Biomech. Eng. 118, 452–463.
183
C OSTA , K.D., H UNTER , P.J., WAYNE , J.S., WALDMAN , L.K., G UCCIONE , J.M., M C C ULLOCH , A.D.
(1996). A three-dimensional finite element method for large elastic deformations of ventricular my-
ocardium: Part II – Prolate spheroidal coordinates. ASME J. Biomech. Eng. 118, 464–472.
DE T OMBE , P.P., TER K EURS , H.E. (1992). An internal viscous element limits unloaded velocity of sarcom-
ere shortening in rat myocardium. J. Physiol. 454, 619–642.
D UCKETT , G., BARKLEY , D. (2000). Modeling the dynamics of cardiac action potentials. Phys. Rev.
Lett. 85 (4), 884–887.
E FIMOV , I.R., E RMENTROUT, B., H UANG , D.T., S ALAMA , G. (1996). Activation and repolarization pat-
terns are governed by different structural characteristics of ventricular myocardium: experimental study
with voltage-sensitive dyes and numerical simulations. J. Cardiovasc. Electrophysiol. 7 (6), 512–530.
E NDRESEN , L.P., S KARLAND , N. (2000). Limit cycle oscillations in pacemaker cells. IEEE Trans. Biomed.
Eng. 47 (8), 1134–1137.
F ENTON , F., K ARMA , A. (1998). Vortex dynamics in three-dimensional continuous myocardium with fiber
rotation – filament instability and fibrillation. Chaos 8 (1), 20–47.
F ITZ H UGH , R. (1961). Impulses and physiological state in theoretical models of nerve membrane. Bio-
phys. J. 1, 445.
F RANZONE , P.C., G UERRI , L., P ENNACCHIO , M., TACCARDI , B. (1998). Spread of excitation in 3-D
models of the anisotropic cardiac tissue. II. Effects of fiber architecture and ventricular geometry. Math.
Biosci. 147 (2), 131–171.
F RANZONE , P.C., G UERRI , L., P ENNACCHIO , M., TACCARDI , B. (2000). Anisotropic mechanisms for mul-
tiphasic unipolar electrograms: simulation studies and experimental recordings. Ann. Biomed. Eng. 28 (11),
1326–1342.
F UNG , Y.C. (1981). Biomechanics: Mechanical Properties of Living Tissues (Springer-Verlag, New York).
G ESELOWITZ , D.B., M ILLER 3 RD , W.T. (1983). A bidomain model for anisotropic cardiac muscle. Ann.
Biomed. Eng. 11 (3-4), 191–206.
G IMA , K., RUDY , Y. (2002). Ionic current basis of electrocardiographic waveforms: a model study. Circ.
Res. 90 (8), 889–896.
G UCCIONE , J.M., M C C ULLOCH , A.D. (1993). Mechanics of active contraction in cardiac muscle: Part I –
Constitutive relations for fiber stress that describe deactivation. J. Biomech. Eng. 115 (1), 72–81.
H ILL , A.V. (1938). Time heart of shortening and the dynamic constants of muscle. Proc. R. Soc. 126, 136–
195.
H ILL , A.V. (1970). First and Last Experiments in Muscle Mechanics (University Press, Cambridge).
H ODGKIN , A.L., H UXLEY , A.F. (1952). A quantitative description of membrane current and its application
to conduction and excitation in nerve. J. Physiol. 117, 500–544.
H OOKS , D.A., T OMLINSON , K.A., M ARSDEN , S.G., L E G RICE , I.J., S MAILL , B.H., P ULLAN , A.J.,
H UNTER , P.J. (2002). Cardiac microstructure: Implications for electrical propagation and defibrillation
in the heart. Circ. Res. 91 (4), 331–338.
H UNTER , P., ROBBINS , P., N OBLE , D. (2002). The IUPS human physiome project. Pflugers Arch. 445 (1),
1–9.
H UNTER , P.J., M C C ULLOCH , A.D., N IELSEN , P.M.F., S MAILL , B.H. (1988). A finite element model of
passive ventricular mechanics. In: Spilker, R.L., Simon, B.R. (eds.), Computational Methods in Bioengi-
neering 9 (ASME, Chicago), pp. 387–397.
H UNTER , P.J., M C C ULLOCH , A.D., TER K EURS , H.E. (1998). Modelling the mechanical properties of
cardiac muscle. Prog. Biophys. Mol. Biol. 69 (2–3), 289–331.
H UNTER , W.C. (1999). Abstract: Making the heart beat: Dynamic models of myocardial sarcomere force
generation. In: 1999 Physiome Symp. in Seattle: Integrated Biology of the Heart, September 9–11.
H UXLEY , A.F. (1957). Muscle structure and theories of contraction. Prog. Biophys. Chem. 7, 255–318.
JAFRI , M.S., D UDYCHA , S.J., O’ROURKE , B. (2001). Cardiac energy metabolism: models of cellular res-
piration. Annu. Rev. Biomed. Eng. 3, 57–81.
K ANAI , A., S ALAMA , G. (1995). Optical mapping reveals that repolarization spreads anisotropically and is
guided by fiber orientation in guinea pig hearts. Circ. Res. 77 (4), 784–802.
K ATZ , A.M. (2001). Physiology of the Heart, third ed. (Lippincott Williams & Wilkins, Philadelphia).
References 185
K EENER , J.P., PANFILOV , A.V. (1997). The effects of geometry and fibre orientation on propagation and
extracellular potentials in myocardium. In: Panfilov, A.V., Holden, A.V. (eds.), Computational Biology of
the Heart (Wiley, New York), pp. 235–258.
K ENTISH , J.C., TER K EURS , H.E., R ICCIARDI , L., B UCX , J.J., N OBLE , M.I. (1986). Comparison between
the sarcomere length-force relations of intact and skinned trabeculae from rat right ventricle. Influence of
calcium concentrations on these relations. Circ. Res. 58 (6), 755–768.
K OGAN , B.Y., K ARPLUS , W.J., B ILLETT , B.S., PANG , A.T., K ARAGUEUZIAN , H.S., K HAN , S.S. (1991).
The simplified Fitzhugh–Nagumo model with action potential duration restitution – Effects on 2D-wave
propagation. Physica D 50 (3), 327–340.
K OHL , P., H UNTER , P., N OBLE , D. (1999). Stretch-induced changes in heart rate and rhythm: clinical ob-
servations, experiments and mathematical models. Prog. Biophys. Mol. Biol. 71 (1), 91–138.
K RUGER , G.W., P OLLACK , J.H. (1975). Myocardial sarcomere dynamics during isometric contraction.
J. Physiol. 51, 627–643.
L ANDESBERG , A., M ARKHASIN , V.S., B EYAR , R., S IDEMAN , S. (1996). Effect of cellular inhomogene-
ity on cardiac tissue mechanics based on intracellular control mechanisms. Am. J. Physiol. Heart Circ.
Physiol. 270 (3 Pt 2), H1101–H1114.
L ANDESBERG , A., S IDEMAN , S. (1994). Coupling calcium binding to troponin C and cross-bridge cycling
in skinned cardiac cells. Am. J. Physiol. Heart Circ. Physiol. 266 (3 Pt 2), H1260–H1271.
L ANGTANGEN , H.P. (1999). Computational Partial Differential Equations Numerical Methods and Diffpack
Programming (Springer-Verlag, Berlin).
L AURITA , K.R., G IROUARD , S.D., ROSENBAUM , D.S. (1996). Modulation of ventricular repolarization
by a premature stimulus Role of epicardial dispersion of repolarization kinetics demonstrated by optical
mapping of the intact guinea pig heart. Circ. Res. 79 (3), 493–503.
L E G RICE , I.J., H UNTER , P.J., S MAILL , B.H. (1997). Laminar structure of the heart: A mathematical model.
Am. J. Physiol. Heart Circ. Physiol. 272 (5 Pt 2), H2466–H2476.
L E G RICE , I.J., S MAILL , B.H., C HAI , L.Z., E DGAR , S.G., G AVIN , J.B., H UNTER , P.J. (1995). Laminar
structure of the heart: ventricular myocyte arrangement and connective tissue architecture in the dog. Am.
J. Physiol. Heart Circ. Physiol. 269 (2 Pt 2), H571–H582.
L E G RICE , I.J., TAKAYAMA , Y., C OVELL , J.W. (1995). Transverse shear along myocardial cleavage planes
provides a mechanism for normal systolic wall thickening. Circ. Res. 77, 182–193.
L IONS , J.L., M AGENES , E. (1972). Non-Homogeneous Boundary Value Problems and Applications, vol. 2
(Springer-Verlag, Berlin).
L UO , C.H., RUDY , Y. (1991). A model of the ventricular cardiac action potential depolarization, repolariza-
tion, and their interaction. Circ. Res. 68 (6), 1501–1526.
L UO , C.H., RUDY , Y. (1994). A dynamic model of the cardiac ventricular action potential. I. Simulations of
ionic currents and concentration changes. Circ. Res. 74 (6), 1071–1096.
M ARON , B.J., B ONOW , R.O., C ANNON , R.O. 3 RD, L EON , M.B., E PSTEIN , S.E. (1987). Hypertrophic
cardiomyopathy. Interrelations of clinical manifestations, pathophysiology, and therapy (1). N. Engl. J.
Med. 316 (13), 780–789.
M C L EAN , M., ROSS , M.A., P ROTHERO , J. (1989). Three-dimensional reconstruction of the myofiber pat-
tern in the fetal and neonatal mouse heart. Anat. Rec. 224 (3), 392–406.
M ICHAILOVA , A.P., S PASSOV , V.Z. (1992). Theoretical model and computer simulation of excitation–
contraction coupling of mammalian cardiac muscle. J. Mol. Cell. Cardiol. 24 (1), 97–104.
M ICHAILOVA , A.P., S PASSOV , V.Z. (1997). Computer simulation of excitation–contraction coupling in car-
diac muscle A study of the regulatory role of calcium binding to troponin C. Gen. Physiol. Biophys. 16 (1),
29–38.
M UZIKANT , A.L., H ENRIQUEZ , C.S. (1997). Paced activation mapping reveals organization of myocardial
fibers: a simulation study. J. Cardiovasc. Electrophysiol. 8 (3), 281–294.
M UZIKANT , A.L., H ENRIQUEZ , C.S. (1998). Validation of three-dimensional conduction models using ex-
perimental mapping: are we getting closer? Prog. Biophys. Mol. Biol. 69 (2–3), 205–223.
N AGUMO , J.S., A RIMOTO , S., Y OSHIZAWA , S. (1962). Proc. IRE 50, 2061.
N EVO , E., L ANIR , Y. (1989). Structural finite deformation model of the left ventricle during diastole and
systole. J. Biomech. Eng. 111 (4), 342–349.
N IELSEN , P.M., L E G RICE , I.J., S MAILL , B.H., H UNTER , P.J. (1991). Mathematical model of geometry
and fibrous structure of the heart. Am. J. Physiol. Heart Circ. Physiol. 260 (4 Pt 2), H1365–H1378.
N OBLE , D. (2001). From genes to whole organs: connecting biochemistry to physiology. Novartis Found
Symp. 239, 111–123; discussion 123–128, 150–159.
N OBLE , D. (2002). Modeling the heart – from genes to cells to the whole organ. Science 295 (5560), 1678–
1682.
O DEN , J.T. (1972). Finite Elements of Nonlinear Continua (McGraw–Hill, New York).
O TANI , N. (2000). Computer modeling in cardiac electrophysiology. J. Comp. Phys. 161, 21–34.
PANERAI , R.B. (1980). A model of cardiac muscle mechanics and energetics. J. Biomech. 13 (11), 929–940.
P RIEBE , L., B EUCKELMANN , D.J. (1998). Simulation study of cellular electric properties in heart failure.
Circ. Res. 82 (11), 1206–1223.
P UGLISI , J.L., B ERS , D.M. (2001). LabHEART: an interactive computer model of rabbit ventricular my-
ocyte ion channels and Ca transport. Am. J. Physiol. Cell. Physiol. 281 (6), C2049–C2060.
Q U , Z., G ARFINKEL , A. (1999). An advanced algorithm for solving partial differential equation in cardiac
conduction. IEEE Trans. Biomed. Eng. 46 (9), 1166–1168.
Q UAN , W., E VANS , S.J., H ASTINGS , H.M. (1998). Efficient integration of a realistic two-dimensional car-
diac tissue model by domain decomposition. IEEE Trans. Biomed. Eng. 45 (3), 372–385.
R ICE , J.J., W INSLOW , R.L., H UNTER , W.C. (1999). Comparison of putative cooperative mechanisms in
cardiac muscle: length dependence and dynamic responses. Am. J. Physiol. Heart Circ. Physiol. 276 (5 Pt
2), H1734–H1754.
ROGERS , J.M., M C C ULLOCH , A.D. (1994). Nonuniform muscle fiber orientation causes spiral wave drift
in a finite element model of cardiac action potential propagation. J. Cardiovasc. Electrophysiol. 5 (6),
496–509.
ROSENBAUM , D.S., K APLAN , D.T., K ANAI , A., JACKSON , L., G ARAN , H., C OHEN , R.J., S ALAMA , G.
(1991). Repolarization inhomogeneities in ventricular myocardium change dynamically with abrupt cycle
length shortening. Circulation 84 (3), 1333–1345.
RUDY , Y. (2000). From genome to physiome: integrative models of cardiac excitation. Ann. Biomed.
Eng. 28 (8), 945–950.
RUEGG , J.C. (1988). Calcium in Muscle Activation: A Comparative Approach (Springer-Verlag, Berlin).
S AFFITZ , J.E., K ANTER , H.L., G REEN , K.G., T OLLEY, T.K., B EYER , E.C. (1994). Tissue-specific deter-
minants of anisotropic conduction velocity in canine atrial and ventricular myocardium. Circ. Res. 74 (6),
1065–1070.
S AXBERG , B.E., G RUMBACH , M.P., C OHEN , R.J. (1985). A time dependent anatomically detailed model
of cardiac conduction. Comput. Cardiol. 12, 401–404.
S ICOURI , S., A NTZELEVITCH , C. (1995). Electrophysiologic characteristics of M cells in the canine left
ventricular free wall. J. Cardiovasc. Electrophysiol. 6 (8), 591–603.
S POTNITZ , H.M., S POTNITZ , W.D., C OTTRELL , T.S., S PIRO , D., S ONNENBLICK , E.H. (1974). Cellular
basis for volume related wall thickness changes in the rat left ventricle. J. Mol. Cell Cardiol. 6 (4), 317–
331.
S TREETER J R ., D.D. (1979). Gross morphology and fiber geometry of the heart. In: Bethesda, M.D. (ed.),
Handbook of Physiology (American Physiological Society), p. 61.
S TREETER J R ., D.D., H ANNA , W.T. (1973). Engineering mechanics for successive states in canine left
ventricular myocardium. I. Cavity and wall geometry. Circ. Res. 33 (6), 639–655.
S TREETER J R ., D.D., S POTNITZ , H.M., PATEL , D.P., ROSS J R ., J., S ONNENBLICK , E.H. (1969). Fiber
orientation in the canine left ventricle during diastole and systole. Circ. Res. 24 (3), 339–347.
S UNDNES , J., L INES , G.T., T VEITO , A. (2001). Efficient solution of ordinary differential equations model-
ing electrical activity in cardiac cells. Math. Biosci. 172 (2), 55–72.
TABER , L.A. (1991). On a nonlinear theory for muscle shells: Part II – Application to the beating left ventri-
cle. J. Biomech. Eng. 113 (1), 63–71.
TAYLOR , T.W., S UGA , H. (1993). Variable crossbridge cycling-ATP coupling accounts for cardiac mecha-
noenergetics. Adv. Exp. Med. Biol. 332, 775–782; discussion 782–783.
TER K EURS , H.E., R IJNSBURGER , W.H., VAN H EUNINGEN , R., N AGELSMIT , M.J. (1980). Tension de-
velopment and sarcomere length in rat cardiac trabeculae. Evidence of length-dependent activation. Circ.
Res. 46 (5), 703–714.
References 187
T ORRENT-G UASP , F. (1973). The Cardiac Muscle (Juan March Foundation, Madrid).
T OZEREN , A. (1985). Continuum rheology of muscle contraction and its application to cardiac contractility.
Biophys. J. 47 (3), 303–309.
T RANUM -J ENSEN , J., W ILDE , A.A., V ERMEULEN , J.T., JANSE , M.J. (1991). Morphology of electrophys-
iologically identified junctions between Purkinje fibers and ventricular muscle in rabbit and pig hearts.
Circ. Res. 69 (2), 429–437.
U SYK , T.P., L E G RICE , I., M C C ULLOCH , A.D. (2002). Computational model of three-dimensional cardiac
electromechanics. Comput. Visual. Sci. 4 (4), 249–257.
U SYK , T.P., M AZHARI , R., M C C ULLOCH , A.D. (2000). Effect of laminar orthotropic myofiber architecture
on regional stress and strain in the canine left ventricle. J. Elasticity 61 (1–3), 143–164.
V ETTER , F.J., M C C ULLOCH , A.D. (1998). Three-dimensional analysis of regional cardiac function:
A model of the rabbit ventricular anatomy. Prog. Biophys. Mol. Biol. 69 (2–3), 157–183.
V ETTER , F.J., M C C ULLOCH , A.D. (2001). Mechanoelectric feedback in a model of the passively inflated
left ventricle. Ann. Biomed. Eng. 29 (5), 414–426.
V ISWANATHAN , P.C., S HAW , R.M., RUDY , Y. (1999). Effects of IKr and IKs heterogeneity on action po-
tential duration and its rate dependence: a simulation study. Circulation 99 (18), 2466–2474.
WALDMAN , L.K., F UNG , Y.C., C OVELL , J.W. (1985). Transmural myocardial deformation in the canine
left ventricle. Normal in vivo three-dimensional finite strains. Circ. Res. 57 (1), 152–163.
W EIDMANN , S. (1970). Electrical constants of trabecular muscle from mammalian heart. J. Physiol. 210 (4),
1041–1054.
W INSLOW , R.L., R ICE , J., JAFRI , S., M ARBAN , E., O’ROURKE , B. (1999). Mechanisms of altered
excitation–contraction coupling in canine tachycardia-induced heart failure, II: Model studies. Circ.
Res. 84 (5), 571–586.
W INSLOW , R.L., S COLLAN , D.F., H OLMES , A., Y UNG , C.K., Z HANG , J., JAFRI , M.S. (2000). Electro-
physiological modeling of cardiac ventricular function: from cell to organ. Annu. Rev. Biomed. Eng. 2,
119–155.
W OLK , R., C OBBE , S.M., H ICKS , M.N., K ANE , K.A. (1999). Functional, structural, and dynamic basis
of electrical heterogeneity in healthy and diseased cardiac muscle: Implications for arrhythmogenesis and
anti-arrhythmic drug therapy. Pharmacol. Ther. 84 (2), 207–231.
W OLK , R., K ANE , K.A., C OBBE , S.M., H ICKS , M.N. (1998). Regional electrophysiological effects of
hypokalaemia, hypomagnesaemia and hyponatraemia in isolated rabbit hearts in normal and ischaemic
conditions. Cardiovasc. Res. 40 (3), 492–501.
W ONG , A.Y. (1971). Mechanics of cardiac muscle, based on huxley’s model: Mathematical stimulation of
isometric contraction. J. Biomech. 4 (6), 529–540.
W ONG , A.Y. (1972). Mechanics of cardiac muscle, based on Huxley’s model: Simulation of active state and
force–velocity relation. J. Biomech. 5 (1), 107–117.
YAN , G.-X., S HIMIZU , W., A NTZELEVITCH , C. (1998). Characteristics and distribution of M cells in arte-
rially perfused canine left ventricular wedge preparations. Circulation 98 (18), 1921–1927.
Y OUNG , A.A., A XEL , L. (1992). Three-dimensional motion and deformation of the heart wall: Estimation
with spatial modulation of magnetization – a model-based approach. Radiology 185 (1), 241–247.
Z YGMUNT , A.C., E DDLESTONE , G.T., T HOMAS , G.P., N ESTERENKO , V.V., A NTZELEVITCH , C. (2001).
Larger late sodium conductance in M cells contributes to electrical heterogeneity in canine ventricle. Am.
J. Physiol. Heart Circ. Physiol. 281 (2), H689–H697.
Mathematical Analysis, Controllability
and Numerical Simulation of a Simple
Model of Avascular Tumor Growth
Jesús Ildefonso Díaz, José Ignacio Tello
Departamento de Matemática Aplicada, Universidad Complutense de Madrid,
Avda Complutense, 28040 Madrid, Spain
Preface
Cancer is one of the most prevalent causes of natural death in the western world, and
a high percentage of people develop some kind of this disease during their lives. For
this reason medicine is one of the scientific fields which found significant interest not
only within the scientific community, but also among the general population. The scien-
tific community comprises medicine, but also other areas of research such us Biology,
Chemistry, Mathematics, Pharmacy or Physics. This is evident from the huge number
of research works and publications in the field and the great quantity of human and
economical resources which have been devoted to cancer research in the last decades.
The development and growth of a tumor is a complicated phenomenon which in-
volves many different aspects from the subcellular scale (gene mutation or secretion of
substances) to the body scale (metastasis). This complexity is reflected by the different
mathematical models given for each phase of the growth. The first phase is known as the
avascular phase, previous to vascularization, and the second one, when angiogenesis
occurs, is known as vascular phase.
The aim of this work is to present the study of the mathematical analysis, the con-
trollability and a numerical simulation for a simple, avascular model of growth of a
tumor. In Section 1, we describe the biological phenomenology of several processes
which influence the growth and development of tumors. The mathematical modelling is
189
190 J.I. Díaz and J.I. Tello
presented by describing different models of partial differential equations (PDE). We fo-

cus our attention on a class of models proposed by G REENSPAN [1972] and B YRNE and
C HAPLAIN [1995], B YRNE [1999a], B YRNE [1999b], B YRNE and C HAPLAIN [1996a],
C HAPLAIN [1996], C HAPLAIN [1999], O RME and C HAPLAIN [1995], T HOMPSON and
B YRNE [1999], WARD and K ING [1998], studied in F RIEDMAN and R EITICH [1999],
C UI and F RIEDMAN [1999], C UI and F RIEDMAN [2000], C UI and F RIEDMAN [2001],
D ÍAZ and T ELLO [2004], D ÍAZ and T ELLO [2003] and by other authors. We prove
the solvability of the model equations and establish uniqueness of solutions under addi-
tional conditions. In Section 6, we study the controllability of the growth of the tumor
by a localized internal action of the inhibitor on a nonnecrotic tumor. It is obvious that
this type of results has merely a mathematical interest and it does not suggest any spe-
cial therapeutical strategy to inhibit tumor growth. Nevertheless our results show that
there is not any obstruction to the controllability (as it appears, for instance, in some
similar PDE’s models: see D ÍAZ and R AMOS [1995]). In a final section, we address the
numerical simulation of the problem.
1. Phenomenology
A tumor originates from mutations of DNA inside cells. In order to create malignant
cells, a sufficiently large number of such mutations has to occur. Factors for mutations
can be external radiation, hereditary causes etc. Eventually, such gene mutations induce
an uncontrolled reproduction, the onset of the formation of a malignant tumor. This
process continues as long as the malignant cells find sufficient supply, and will generate
a small spheroid of a few millimeters. During this time, called the avascular phase,
nutrients (glucose and oxygen) arrive at the cells through diffusion. As the spheroid
grows, the level of nutrients in the interior of the tumor decreases due to consumption by
the outer cells. When the level of concentration of nutrients in the interior falls bellow
a critical level, the cells cannot survive, a phenomenon called necrosis, and an inner
region is formed in the center of the tumor by the dead cells, which decompose into
simpler chemical compounds (mainly water). At this time, one can distinguish several
regions in the tumor: a necrotic region in the center, an outer region, where mitosis
(division of cells) occurs, and a region in between where the level of nutrients suffices
for the cells to live, but not to proliferate. Until this moment, the tumor is a multicell
spheroid whose radius is no more than a few millimeters.
The cells of the tumor secrete some chemical substances, known as Tumor Angio-
genesis Factors (TAFs). These substances diffuse through the surrounding tissue. TAFs
stimulate endothelial cells (ECs), located in neighboring blood vessels. Endothelial cells
are thin cells which form the basement membrane of the blood vessels. When ECs are
stimulated by TAFs, they destroy the membrane basement (by secretion of proteases
and collagenases) and migrate towards the tumor forming capillary sprouts. These grow
thanks to the proliferation of ECs and other substances located in the extracellular ma-
trix (as fibronectin), forming a capillary network. Initially, the ECs move forming par-
allel vessels and as sprouts are closer to the tumor, the sprouts branch out and connect.
This process of formation of new vessels, known as angiogenesis, is one of the most
decisive steps in the growth of a tumor. Angiogenesis is present in other contexts of life,
as well, like in wound healing or in the formation of embryos.
Mathematical analysis, controllability and numerical simulation 191
The connection of the blood vessels to the tumor supplies nutrients to the malignant
cells, aiding a faster proliferation of the tumor’s cells. This phase of the tumor is known
as vasculature phase and is characterized by an aggressive growth.
Finally, the cells of the tumor invade the surrounding tissue and metastasizing to other
parts of the body. The circulatory and lymphatic systems are used by the malignant cells
for transport to another sites. The process in which cells leave the tumor and enter into
the vessels is known as intravasation. Cancer cells, which survive in the blood flow and
escape from the circulatory system, arrive at a new site, where a new colony of cells
may grow. Fortunately, less than 0.05 per cent of cells which were introduced in the
circulation are able to create new colonies. Each tumor has a preference to metastasize
to a specific organ.
During the growth of a tumor, the immune system competes with the malignant cells;
it will be activated through the recognition of the cancer cells by the immune cells.
Macrophages (Ms) are a type of white blood cells, which migrate into the tumor to the
regions with low oxygen (hypoxic regions) in the interior of the tumor through the ex-
ternal layer of well nourished cells of the tumor. Ms move to the tumor (by chemotaxis)
attracted by macrophage chemoattractants, which is secreted by the tumor. A cytotoxic
substance is secreted into the tumor’s cell which kills it. Ms may also help the growth
of the tumor secreting other chemical substances which help angiogenesis.
It is the main strategy of all cancer therapies (apart from surgery) to inhibit the growth
of tumors with tools adapted to the phase the tumor is in. E.g., chemotherapy or radiation
therapy are intended to destroy cells of tumor, other treatments try to stimulate cells of
the immune system. The first type of therapy is nonselective, destroying both, malignant
cells and cells of the immune system. Another therapy based on genetic engineering is
being studied. The idea is to insert a therapeutic gene into the cells of a patient and
re-inject them back into the patient.
2. Mathematical modelling
Mathematical modelling of the growth of a tumor have been studied by several authors
during the last thirty years in many different works.
Among the many different PDE models we can introduce (following F RIEDMAN
[2002]) a rough classification into two classes: the mixed models, in which all the dif-
ferent population of cells are continuously present everywhere in the tumor, at all the
times, and segregated models, perhaps less realistic but relevant for in vitro experiments,
in which the different populations of cells are separated by unknown interfaces or free
boundaries. Our analysis will be restricted to the second class of models (some refer-
ences on mixed models can be found in B ELLOMO and P REZIOSI [2000], DE A NGE -
LIS and P REZIOSI [2000], C HAPLAIN and P REZIOSI [2002] and F RIEDMAN [2002]).
Moreover, we shall consider spherical tumors (for other free boundary type tumors,
without symmetrical shape, arising in tumoral masses growing around a blood vessel
see, e.g., B ERTUZZI , FASANO , G ANDOLFI and M ARANGI [2002] or BAZALIY and
F RIEDMAN [2003]).
In this section, we describe different mathematical models for each phase. A first and
simple model describing the avascular phase was presented in G REENSPAN [1972], as-
suming spherical symmetry in R3 . The outer boundary delimiting the tumor is denoted
by R(t) and the concentration of nutrients and inhibitors by σ and β, respectively. Ac-
cording to principle of conservation of mass, the tumor mass is proportional to its vol-
ume 43 πR 3 (t), assuming the density of the cell mass is constant. The balance between
the birth and death rate of cells is given as a function of the concentration of nutrients
and inhibitors. Let S be this balance, then after normalizing, we obtain the law

d 4
πR 3 (t) =
S σ (x̃, t), β(x̃, t) dx̃.
dt 3 {|x̃|<R(t )}
Depending on the author, the function S can be written in different ways. G REENSPAN
[1972] studied the problem in the presence of an inhibitor, and the possibility that this af-
fects mitosis, when the concentration of the inhibitor is greater than a critical level β̃. He
proposed S(σ, β) = sH (σ − σ̃ )H (β̃ − β), where H (·) denotes the maximal monotone
graph of R2 associate with the Heaviside function, i.e., H (k) = 0 if k < 0, H (k) = 1 if
k > 0 and H (0) = [0, 1]. B YRNE and C HAPLAIN [1996a] study the growth when the
inhibitor affects the cell proliferation and propose S(σ, β) = s(σ − σ̃ )(β̃ − β) (for a
positive constant s). In the absence of inhibitors or in case that the inhibitor does not
affect mitosis, they choose S(σ, β) = sσ (σ − σ̃ ). F RIEDMAN and R EITICH [1999] and
C UI and F RIEDMAN [2000] study the asymptotic behavior of the radius, R(t), with the
cell proliferation rate free of the action of inhibitors. They assume that S = s(σ − σ̃ ),
where sσ is the cell birth-rate and the death-rate is given by s σ̃ (see also the survey
S LEEMAN [1996]).
We assume that the tumor is composed of an homogeneous tissue and that the dis-
tribution of the concentration of nutrients σ is governed by a PDE in the spheroid. As-
suming that there is no inhibitor, that the tumor has not necrotic core and that diffusion
is high, we obtain the equation
d1 σ = λσ, |x| < R,

where λσ represents the nutrient consumption by cells and d1 is the diffusion coeffi-
cient.
In necrotic tumors, an inner free boundary appears, which is denoted by ρ(t). It
separates the necrotic core (where σ falls below σn ) from the remaining part. A model
for necrotic tumors was presented in B YRNE [1997a], who proposes the equation

0 = σ − λH |x| − ρ(t) , |x| < R(t),
where the effect of time-delay appears in the radial growth. In addition, asymptotic
techniques are used to show the effect of the delay terms.
Several authors (A DAM [1986] and B RITTON and C HAPLAIN [1993]) studied a
model proposed by S HYMKO and G LASS [1976] where cell proliferation is controlled
by chemical substances Growth inhibitor factor (GIFs) as chalones. GIFs secreted by
cells reduce the mitotic activity. Two different kinds of inhibitors appear, depending on
the phase of the cell cycle stage at which inhibition occurs. The inhibitor can act before
DNA synthesis (as epidermal chalon in Melanoma or granulocyte chalon in Leukemia)
or before mitosis (see ATTALLAH [1976]). The concentration of GIF (denoted by C) is
modeled by one PDE in a bounded domain Ω of R3 ,

∂C
= dC + f (C) + S(x), x ∈ Ω, t > 0, (2.1)
∂t
∂C
D + P C = 0, x ∈ ∂Ω, t > 0, P 0, (2.2)
∂n
C(x, t) = C0 (x), x ∈ Ω, (2.3)
where S(x) is a source term and f (C) represents the decay of GIF (see A DAM and
B ELLOMO [1997]).
In 1972, G REENSPAN [1972] proposed a radially symmetric model employing the
Heaviside function H for modelling the necrotic part. The avascular model considers a
chemical inhibitor β, which is produced in the necrotic core. The distribution of nutri-
ents σ̂ is given by the equation
∂σ
− d1 σ = −λ(σB − σ )H |x| − ρ H R − |x| , (2.4)
∂t
where R is the outer boundary of the tumor and ρ is the radius of the necrotic core.
The chemical substance “β” (produced within the tumor) inhibits the mitosis of can-
cer cells without causing their death and satisfies the diffusion equation
∂β
− d2 β = P H |x| − ρ H R − |x| − Pd H ρ − |x| . (2.5)
∂t
This model, proposed by Greenspan, has been studied by several authors in the last
thirty years. We shall focus on the study of a similar model and detail the modelling and
some mathematical results in the next section.
When asymmetric distribution of nutrients or displacement of cells produced by
nonuniform density appears in the interior of the spheroid tumor, the internal forces
may break the symmetry of the outer boundary. Several authors have studied, in dif-
ferent models, the symmetry breaking of the boundary. G REENSPAN [1976] studied a
model where the pressure p of the cancer cells satisfies
p = S,
inside the tumor, where S is the rate of volume lost per unit volume (assumed con-
stant). The distribution of nutrients σ satisfies a elliptic equation outside of the tumor.
Using Darcy’s law, (the velocity v of the boundary is proportional to the gradient of p)
that is v = µ∇p, with suitable boundary conditions for p and σ , Greenspan obtains
nonsymmetric explicit solutions using spherical harmonics.
Darcy’s law has been used in different models in order to describe the movement of
the free boundary. B YRNE [1997b], B YRNE and C HAPLAIN [1996b] and B YRNE and
M ATTHEWS [2002] propose similar models improving G REENSPAN [1976]; they study
the stability of radially symmetric solutions via perturbations with spherical harmon-
ics. F RIEDMAN and R EITICH [2001] study the bifurcation of non-symmetric solutions
from any radially symmetric steady state. Bessel functions are used in F RIEDMAN and
R EITICH [2001] and also in F RIEDMAN , H U and V ELÁZQUEZ [2001] in a protocell
model.
L EVINE , S LEEMAN and N ILSEN -H AMILTON [2000] and L EVINE , PAMUK , S LEE -
MAN and N ILSEN -H AMILTON [in press] (see also H OLMES and S LEEMAN [2000]) de-
veloped models of angiogenesis based on analysis of the relevant biochemical processes
and on the methodology of the reinforced random walk of OTHMER and S TEVENS
[1997]. A mathematical analysis of the model proposed in L EVINE , S LEEMAN and
N ILSEN -H AMILTON [2000] have been performed in F ONTELOS , F RIEDMAN and
H U [2002]. Their model involves several diffusing populations and several chemical
species. Another model of angiogenesis with one diffusing population and two non-
diffusing ones, was developed in A NDERSON and C HAPLAIN [1998] and C HAPLAIN
and A NDERSON [1997]. They denote the density of the endothelial cells by p, the con-
centration of the tumor angiogenesis factor (secreted by the tumor) by c, and w repre-
sents the density of the fibronectin cells, then

∂p α ∂w
= div ∇p − p ∇c + ρ∇w , = γp(1 − w),
∂t 1+c ∂t
∂c
= −µpc,
∂t
where α, ρ, γ and µ are positives constants. The asymptotic behavior of the solutions
has been studied for some values of the parameters and special initial data in F RIEDMAN
and T ELLO [2002]. A computational approach is used by VALENCIANO and C HAPLAIN
[2003a], VALENCIANO and C HAPLAIN [2003b] to obtain numerical solutions for simi-
lar models. L EVINE and S LEEMAN [1997] study the chemotaxis equations developed in
the context of reinforced random walks. They use the classification of the second order
part of a modified equation in the “Hodograph plane” and study the existence of blow
up of solutions in finite time.
Recently, B ERTUZZI , FASANO , G ANDOLFI and M ARANGI [2002] have developed a
model for the phase transition in tumor cells and their migration towards the periphery.
The macrophages cells are part of the response of the immune system to cancer; their
movement has been modeled by different authors (see OWEN and S HERRATT [1999]).
3. A simple mathematical model
In this section we describe a simple mathematical model which will be studied through-
out the remainder of this work. It belongs to a group of first generation cancer models
with Greenspan’s model (2.4), (2.5) being one of the earliest ones. Similar models have
been proposed and studied by several authors (B YRNE and C HAPLAIN [1996a], F RIED -
MAN and R EITICH [1999], C UI and F RIEDMAN [2000], C UI and F RIEDMAN [2001]
and D ÍAZ and T ELLO [2004], D ÍAZ and T ELLO [2003]). We assume that the density of
live cells is proportional to the concentrations of the nutrients σ . The tumor occupies a
ball in R3 of radius R(t) which is unknown (which is reason why R is usually called
the free boundary of the problem).
The tumor comprised a central necrotic core of dead cells, the necrotic core is covered
with a layer (of living cells) resulting in a second free boundary denoted by ρ(t) in
G REENSPAN [1972].
The transfer of nutrients to the tumor through the vasculature occurs below a certain
level σB , and it is done with a rate r1 . During the development of the tumor, the immune
system secretes inhibitors as a immune response to the foreign body. The structure of
inhibitor absorption is similar to the transference of nutrients (for a constant r2 ). If
we assume that the nutrient consumption rate is proportional to the concentrations of
nutrients, the nutrient consumption rate is given by λσ . Both processes, consumption
and transference, occur simultaneously in the exterior of the necrotic core, where cells
are inhibited by β̂. We assume that the host tissue is homogeneous and that the diffusion
coefficient, d1 , is constant. The reaction between nutrients and inhibitors can be globally
modelled by introducing the Heaviside maximal monotone graph (as function of σ̂ ) and
some continuous functions gi (σ̂ , β̂). Then σ̂ satisfies
∂σ
− d1 σ ∈ r̂1 (σB − σ ) − λ1 σ − λβ H (σ − σn ) + ĝ1 (σ, β). (3.1)
∂t
We also assume a constant diffusion coefficient for the inhibitor concentration β̂, d2 .
The model considers the permanent supply of inhibitors, modeled by f˜ and localized
on a small region ω0 inside the tumor. This term f˜ was introduced in D ÍAZ and T ELLO
[2003] to control the growth of the tumor. Then β satisfies
∂β
− d2 β ∈ −r2 (β − βB )H (σ − σn ) + ĝ2 (σ, β) + f˜χω0 , (3.2)
∂t
adding initial and boundary conditions, we obtain
σ (x̃, t) = σ , β(x̃, t) = β, |x̃| = R(t), (3.3)

σ (x̃, 0) = σ0 (x̃), β(x̃, 0) = β0 (x̃), |x̃| < R0 . (3.4)
In this formulation, the presence of the maximal monotone graph H is the reason
why the symbol ∈ appears in Eq. (3.2) instead of the equal sign (a precise notion of
weak solution will be presented later). Different constants appears in the equations and
boundary conditions which lead to a wide variety of special cases: σn is the level of
concentration of nutrients above which the cells can live (below this level the cells die
by necrosis), σ and β are the concentration of nutrients and inhibitors in the exterior of
the tumor. The diffusion operator is the Laplacian operator and χω0 denotes the char-
acteristic function of the set ω0 (i.e., χω0 (x̃) = 1, if x̃ ∈ ω0 , and χω0 (x̃) = 0, otherwise).
Notice that the above formulation is of global nature and that the inner free boundary
ρ(t) is defined implicitly as the boundary of the set {r ∈ [0, R(t)): σ σn }. So, if
for instance, the initial datum σ0 satisfies σ0 (x̃) = σn on [0, ρ0 ], for some ρ0 > 0 and
ĝ1 (σn , β) ∈ [0, r1 (σB − σn ) − λσn ] for any β 0, the above formulation leads to the
associate double free boundary formulation in which σ̂ satisfies

 ∂σ

 − d1 σ + λ1 σ = r̂1 (σB − σ ) + ĝ1 (σ, β), ρ(t) < |x̃| < R(t),


 ∂t
σ (x̃, t) = σn , |x̃| ρ(t),



 σ (x̃, t) = σ , |x̃| = R(t),


R(0) = R0 , ρ(0) = ρ0 , σ (x̃, 0) = σ0 (x̃), ρ0 < |x̃| < R0 .
The free boundary R(t) is described by the ODE presented in Section 2,

d 4
πR 3 (t) =
S σ (x̃, t), β(x̃, t) dx̃, R(0) = R0 . (3.5)
dt 3 {|x̃|<R(t )}
4. Existence of solutions
In this section, we study the existence of solutions to (3.1)–(3.5) after introducing some
structural assumptions on ĝi and S. We also introduce some functional spaces and a
useful change of variables. The existence result is presented in Theorem 4.1 and proved
by using a Galerkin approximation based on a weak formulation of the problem.
We shall assume that the reaction terms ĝi and the mass balance of the tumor S
satisfy:

ĝi are piecewise continuous, ĝi (a, b) c0 + c1 |a| + |b| , (4.1)
2

S is continuous and − λ0 S(a, b) c0 + c1 |a| + |b| 2
(4.2)
for some positives constants λ0 , c0 , c1 .
The above assumptions ((4.1) and (4.2)) do not constitute biological restrictions, and
previous models satisfy them provided σ and β are bounded. They are introduced in or-
der to carry out the mathematical treatment, and its great generality allows us to handle
all the special cases from the literature previously mentioned. They are relevant due to
its generality. It is possible to show that the absence of one (or both) of the conditions
implies the occurrence of very complicated mathematical pathologies, and much more
sophisticated approaches would be needed for proving that the model admits a solution
(in some very delicate sense).
We introduce the change of variables,
x̃
x = (x1 , x2 , x3 ) = , (4.3)
R(t)

u(x, t) = σ R(t)x, t − σ (4.4)
and

v(x, t) = β R(t)x, t − β. (4.5)
Let the unit ball {x ∈ R3 : |x| < 1} be denoted by B and define functions from R2 to
2R
2
by

g1 σ − σ , β − β := r̂1 (σB − σ ) − λ1 σ − λβ H (σ − σn ) + ĝ1 (σ, β),
g σ − σ , β − β := −r (β − β )H (σ − σ ) + ĝ (σ, β),
(4.6)
2 2 B n 2
4
S σ − σ , β − β := S(σ, β) (4.7)
3π
and

f (x, t) := f˜ xR(t), t ,
ω0t = (x, t) ∈ B × [0, T ]: R(t)x ∈ ω0 .
Problem (3.1)–(3.5) becomes


 ∂u d1 R (t)

 − u − x · ∇u ∈ g1 (u, v), x ∈ B, t > 0,

 ∂t R(t) 2 R(t)



 R (t)

 ∂v d2
 − v − x · ∇v ∈ g2 (u, v) + f χ
ω0t , x ∈ B, t > 0,
∂t R(t)2 R(t)
(4.8)

 dR(t)

 R(t)−1 = B S(u, v) dx, t > 0,

 dt



 u(x, t) = v(x, t) = 0, x ∈ ∂B, t > 0,


R(0) = R0 , u(x, 0) = u0 (x), v(x, 0) = v0 (x), x ∈ B.
We introduce the Hilbert spaces
H(B) := L2 (B)2 , V(B) = H01 (B)2

and define inner products by

Φ, Ψ H(B) = Φ · Ψ t dx, Φ, Ψ V(B) = di (∇Φi )t · ∇Ψi dx
B i=1,2 B
for all Φ = (Φ1 , Φ2 ), Ψ = (Ψ1 , Ψ2 ).

For the sake of notational simplicity we use H = H(B) and V = V(B). Given T > 0,
we introduce U = (u, v), U0 = (u0 , v0 ) and define G : R2 → 2R × 2R and F : (0, T ) ×
2 2
B → R2 by

G(U ) = g1 (u, v), g2 (u, v) , F (t, x) = 0, f (t, x)χ
ωt . 0
We have

G(U ) = g1 (u, v) + g2 (u, v) C0 + C1 |U | = C0 + C1 |u| + |v| . (4.9)
D EFINITION . (U, R) ∈ L2 (0, T : V) × W 1,∞ (0, T : R) is a weak solution of the prob-

lem (4.8) if there exists g ∗ = (g1∗ , g2∗ ) ∈ L2 (0, T : H) with g ∗ (x, t) ∈ G(U (x, t)) a.e.
(x, t) ∈ B × (0, T ) satisfying
T T T
−U, Φt H dt + ã(t, U, Φ) dt = g ∗ , ΦH dt
0 0 0
T
+ U0 , Φ(0)H + F (t), ΦH dt,
0
∀Φ ∈ L2 (0, T : V) ∩ H 1 (0, T : H) with Φ(T ) = 0, where

1 R (t)
ã(t, U, Φ) := U, ΦV − x · ∇U, ΦH (4.10)
R 2 (t) R(t)
and R(t) is strictly positive and given by

dR(t)
R(t)−1 = S U (x, t) dx for t ∈ (0, T ).
dt B
D EFINITION . (σ, β, R) is a weak solution of (3.1)–(3.5) if

x̃ x̃
σ (x̃, t) = u , t + σ and β(x̃, t) = v , t + β,
R(t) R(t)
for t ∈ (0, T ) and x̃ ∈ R3 , |x̃| R(t), where (U = (u, v), R) is a weak solution of (4.8)
for any T > 0.
R EMARK 4.1. The definition of weak solution and the structural assumptions on G
imply that ∂U /∂t ∈ L2 (0, T : V(B) ) and the equation holds in D (B × (0, T )).
T HEOREM 4.1. Assume (4.1), (4.2), R0 > 0 and σ0 , β0 ∈ L2 (0, R0 ), then problem
(3.1)–(3.5) has at least a weak solution for each T > 0.
P ROOF. We shall use a Galerkin method to construct a weak solution. Let R(t) ∈
W 1,∞ (0, T : R) such that R (t)/R(t) −λ0 a.e. t ∈ (0, T ). For fixed t ∈ (0, T ), we
consider the operator A(t) ≡ A(R(t)) : V → V defined by

d1
− R(t )2
u − RR(t(t)) x · ∇u 0
A R(t) (U ) = .
0 d2
− R(t )2
v − RR(t(t)) x · ∇v
A(t) defines a continuous, bilinear form on V × V,
ã(t : ·, ·) : V × V → R
for a.e. t ∈ (0, T ) (see (4.10)). Since R (t)/R(t) −λ0 , ã satisfies
1 R (t)
ã(t, U, U ) = U, U V −
x · ∇U, U H
R 2 (t) R(t)
1 R (t)
= 2 U, U V + U, U H
R (t) 2R(t)
−2 λ0
max R(t) U 2V − U 2H .
0<t <T 2
Now we establish some a priori estimates which will be used later. In fact, those
estimates can be applied even for other existence methods, different from the Galerkin-
type one, as, for instance, iterative methods, fixed point methods, etc. (see, for instance,
S HOWALTER [1996]).
L EMMA 4.1.

U 2H C02 exp (λ0 + 2C1 + 1)T − 1 + F 2L2 (0,T :H) + U0 2H .
P ROOF. Inserting U t as test function into the weak formulation of (4.8), one obtains

d 1 2 ∗
U dx + ã(t, U, U ) + g (U )U dx = F · U t dx
t
dt B 2 B B
for some g ∗ ∈ L2 ((0, T ) × B)2 and g ∗ (x, t) ∈ G(U (x, t)) for a.e. (x, t) ∈ B × (0, T ).
The definition of ã yields
1 d λ0
U 2H − U 2H g ∗ H + F H U H . (4.11)
2 dt 2
Thus by Young’s inequality and (4.9) imply

1 d λ0 1 1
U H −
2
+ C1 + U 2H C02 + F 2H .
2 dt 2 2 2
Integrating with respect to time, we get

1 1 λ0 1 1
U 2H − U0 2H − + C1 + U 2L2 (0,T :H) C02 T + F 2L2 (0,T :H)
2 2 2 2 2
and by Gronwall’s lemma,

U 2H C02 exp (λ0 + 2C1 + 1)T − 1 + F 2L2 (0,T :H) + U0 2H C. (4.12)

R EMARK 4.2. Since U is bounded in H (by (4.12)), R satisfies

t 1
R(t) = R0 exp S(U ) dx dt R0 eK1 t (4.13)
0 0
and
R(t) R0 exp{−λ0 t}, (4.14)
consequently, R ∈ W 1,∞ (0, T ).
L EMMA 4.2. U L2 (0,T :V) K(T , F, G, U0 ).
P ROOF. Selecting U as test function in (4.8), we have

D λ0
U 2L2 (0,T :V) − U 2L2 (0,T :H)
R02 e2K1 T 2

C1 U 2L2 (0,T :H) + C0 + F L2 (0,T :H) U L2 (0,T :H) .
By (4.12), we get
U L2 (0,T :V ) K(F, G, U0 , T ). (4.15)

R EMARK 4.3. By Lemma 4.2 and Remark 4.2, we get that

d1 d2
ut − 2 u ∈ L2 0, T : L2 (B) , vt − 2 v ∈ L2 0, T : L2 (B)
R R
and obtain the extra regularity
2
Ut , U ∈ L2 0, T : L2 (B) . (4.16)
Now, as previously in the proof of Theorem 4.1, we consider the approximate prob-
lem

 ∂U ε

 + A R ε (t) U ε = Gε U ε + F (t) on B × (0, T ),

 ∂t
U ε (0, x) = U0 , U ε = 0 on ∂B, (4.17)



 1 dR ε
 = S(U ε ) dx,
R ε dt B
where Gε = (g1ε , g2ε ) is a Lipschitz continuous function such that
Gε → G when ε → 0 a.e. in R2 .
Gε is obtained replacing H by
 0 if s < 0,


s 1
if 0 s ,
H (s) = ε
ε
ε


 1 if s > 1 .
ε
Now, we apply the Galerkin method to the approximated problem. Let λn and φn ∈
H01 (B) for n ∈ N be the eigenvalues and eigenfunctions associated to − satisfying
−φn = λn φn .
We consider Vm the finite-dimensional vector space spanned by {φ1 , . . . , φm }. We
search for a solution Umε ∈ L2 (0, T : Vm ) of the problem

 d ε ε ε

 Um + A Rm (t) Um = Gε Umε + Fm (t),


 dt
Umε (0) = U0,m
ε
, (4.18)



 ε

 Rmε (t)−1 dRm (t) = S Umε (x, t) dx,
dt B
ε = P (U ) (where P is the orthogonal projection
where the initial conditions U0,m m 0 m
from L2 (B) onto Vm ) and Fm = Pm (F ). Then
t

Rmε
(t) = R0 exp S Umε (x, s) dx ds .
0 B
P ROPOSITION 4.1. (4.18) has a unique solution Umε for any T < ∞.
P ROOF. Problem (4.18) can be written as a suitable nonlinear ordinary differential sys-
tem. Let Umε = (uεm , vm
ε ) be defined by

uεm (t) = anεm (t)φn , vmε
(t) = bnεm (t)φn
n=1,...,m n=1,...,m
and denote

a εm = a1εm , a2εm , . . . , am
εm
, bεm = b1εm , b2εm , . . . , bm
εm
,

λa = λ1 a1 , . . . , λm am
εm εm
and λb = (λ1 b1 , . . . , λm bm ).
εm εm
Then a εm , b εm and Rm
ε satisfy
λa εm εm
ȧ εm + ε 2
+ φε a εm , bεm Lm
1 a ,b = g1m a εm , bεm ,
(Rm )
λb εm εm
ḃεm + ε 2 + φε a εm , bεm Lm 2 a ,b = g2m a εm , bεm + F m (t),
(Rm )
Ṙmε
ε
= φε a εm , bεm ,
Rm
where

φε a εm , bεm = S Umε dx,
B
εm εm
Lm
1 a ,b = x · ∇uεm φn dx for n = 1, . . . , m,
B
εm εm
Lm
2 a ,b = x · ∇vm
ε
φn dx for n = 1, . . . , m,
B

g1m a εm , bεm = g1ε (uεm , vm
ε
)φn dx for n = 1, . . . , m,
B

g2m a εm , bεm = g2ε uεm , vm
ε
φn dx for n = 1, . . . , m.
B
Since Gε is a Lipschitz function, we obtain that there exists a unique solution
a εm , bεm , R εm to the system for T small enough. Moreover, (4.12) and (4.14) hold,
and we get the existence of a solution of (4.18) for any T < ∞. By (4.15) and (4.16),
{(Umε , dtd Umε )}m=1,∞ is uniformly bounded in L2 (0, T : V) × L2 (0, T : V ). So, there
exists a subsequence Umi ε ∈ L2 (0, T : V) with d U ε ∈ L2 (0, T : V ) such that
dt mi

d ε d
Umi ε
, Umi U ε, U ε weakly in L2 (0, T : V) × L2 (0, T : V ),
dt dt
ε R ε weakly in W 1,p (0, T ) for p < ∞. Taking limits when mi → ∞, we get
and Rmi
the existence of a weak solution to (4.17) for any T < ∞.
To end the proof of Theorem 4.1, we take limits in the equation when ε → 0. We
employ (4.12) and (4.14) and the compact embedding H10 (B) ⊂ Ls (B) (for s < 6) in
order to obtain the existence of a subsequence U εi such that
2
U εi → U in L2 0, T : Ls (B)
and in particular
U εi → U in L2 (0, T : H)
(see, e.g., S IMON [1987]). Since

H ε uε + σ h ∈ H (u + c) weakly in L2 0, T : Ls (B)
and

vε → v in L2 0, T : Ls (B)
(see Lemma 3.4.1 of V RABIE [1995]), we have

Gεi U εi g ∗ ∈ G(U ) weakly in L1 (0, T : H).
Since |R | C, there exists a subsequence Rεij such that
Rεij R weakly in W 1,p (0, T ), p < ∞,

and we deduce that Rεij → R in C 0 ([0, T ]). Finally, taking limits in the weak formula-
tion of the problem (4.17), we get

T T T T
Ut , ΦH dt + ã R(t), U, Φ dt + g ∗ , ΦH dt = F, ΦH dt
0 0 0 0
for all Φ ∈ L2 (0, T : V ) and, moreover,

−1 dR(t)

R(t) = S U (x, t) dx.
dt B
Notice that
T R T
Rεij
εij
x · ∇uεij ψ dx dt = uεij ψ − uεij x · ∇ψ dx dt
0 Rεij B 0 Rεij B
and
T
Rεij T
Rεij
x · ∇vεij ψ dx dt = vεij ψ − vεij x · ∇ψ dx dt.
0 Rεij B 0 Rεij B
We conclude that (σ, β, R) defined by

x̃ x̃
σ (t, x̃) = u t, + σ and β(t, x̃) = v t, +β
R(t) R(t)
is a weak solution to (3.1)–(3.5). The additional regularity

σt − d1 σ and βt − d2 β ∈ L 2
0, R(t) × {t}
t ∈[0,T ]
follows from the fact that

∂U
(t) + A R(t) U (t) ∈ L2 0, T : L2 (B)2 .
∂t
5. Uniqueness of solutions
We begin by pointing out that if, for instance,

r1 σB
σn , r1 σB > 0, ĝ1 (σ̂ , β̂)
r1 + λ
is a decreasing function of σ̂ and independent of β̂ and the initial datum σ0 (x̃) is such
that σ0 (ρ0 ) = σ0 (ρ0 ) = 0, then it is possible to adapt the arguments of D ÍAZ and T ELLO
[1999] in order to construct more than one solution of problem (3.1)–(3.5). This and the
presence of non-Lipschitz terms at both equations clarify that any possible uniqueness
result will require an significant set of additional conditions.
In this section we prove the uniqueness of solution for two different cases. C UI and
F RIEDMAN [2000] prove uniqueness of radial symmetric solutions without forcing term
(i.e., f = 0).
5.1. 3-dimensional case with forcing term
When a tumor does not have a necrotic core, Eqs. (3.1) and (3.2) simplify such that
reaction terms become linear, i.e., the nutrients concentration σ̂ and the inhibitors con-
centration β̂ satisfy
∂ σ̂
− d1 σ̂ − r̂1 (σB − σ̂ ) + λ1 σ̂ + λβ̂ = 0, |x| < R(t), t ∈ (0, T ),
∂t
∂ β̂
− d2 β̂ − r2 (βB − β̂) = f χω0 , |x| < R(t), t ∈ (0, T ).
∂t
For notational convenience we shall assume that the diffusion coefficients d1 and d2
are equal and constant d1 = d2 = d. Thus by normalizing the unknown densities
r̂1 σB + λβB
σ := σ̂ − , β := β̂ − βB ,
(r̂1 + λ1 )
and setting
3
r1 := r̂1 + λ1 , S(σ, β) := S(σ̂ , β̂),
4π
we arrive at the formulation
∂σ
− dσ + r1 σ + λβ = 0, |x| < R(t), t ∈ (0, T ), (5.1)
∂t
∂β
− dβ + r2 β = f χω0 , |x| < R(t), t ∈ (0, T ), (5.2)
∂t
2 dR(t)
R(t) = S(σ, β) dx, R(0) = R0 , t ∈ (0, T ), (5.3)
dt |x|<R(t )
σ (x, 0) = σ0 (x), β(x, 0) = β0 (x), |x| < R0 , (5.4)
σ (x, t) = σ , β(x, t) = β, |x| = R(t), t ∈ (0, T ), (5.5)
where R0 > 0, the normalized nutrient and inhibitor densities at the exterior of the
tumor σ , β and the initial densities (σ0 , β0 ) are known. We introduce again the changes
of unknown and variables (4.3)–(4.5) and set
t
˜t (t) := R −2 (ρ) dρ. (5.6)
0
Note that since R is a continuous function and 1/R 2 (t) > 0, we obtain that t˜(t) ∈
]) and employing the implicit function theorem, one derives the existence of
C 1 ([0, T
the inverse function t (t˜) ∈ C 1 ([0, T ]). Then, problem (5.1)–(5.5) reduces to
∂u
+ A(u) + R 2 r1 u = R 2 r1 σ + λ v + β , x̃ ∈ B, t˜ ∈ (0, T ), (5.7)
∂ t˜
∂v ),
+ A(v) + R 2 r2 v = R 2 f χω0t˜
− R 2 r2 β, x̃ ∈ B, t˜ ∈ (0, T (5.8)
∂ t˜
d
R(t˜) R(t˜) = S u(x̃, t˜) + σ , v(x̃, t˜) + β dx̃, R(0) = R0 , (5.9)
dt˜ B
u(x̃, t˜) = v(x̃, t˜) = 0, x̃ ∈ ∂B, t˜ ∈ (0, T), (5.10)
u(x̃, 0) = u0 (x̃) = σ0 (x̃R0 ), v(x̃, 0) = v0 (x̃) = β0 (x̃R0 ), (5.11)
where T ω0t˜ = {x̃ ∈ B: R(t (t˜))x̃ ∈ ω0 }, for any t˜ ∈ [0, T
= t˜(T ), ] and
A(w) := −dw − R Ṙ x̃ · ∇w.

We assume that

S ∈ W 1,∞ R2 , (5.12)

f χ
ω0t˜
∈ Lp (0, T ) × Ω , p > 4, (5.13)
2
(σ0 , β0 ) ∈ W 2,∞ B(R0 ) . (5.14)
L EMMA 5.1. Assume (5.12)–(5.14), then the solution (u, v, R) to the problem (5.7)–
(5.11) satisfies

u ∈ Lq 0, T : W 2,q (B) ∩ W 1,q 0, T
: Lq (B)
for all 1 < q < ∞ and

v ∈ Lp 0, T : W 2,p (B) ∩ W 1,p 0, T : Lp (B) .
P ROOF. By Theorem 4.1, we know that

2
(u, v, R) ∈ L2 0, T : H 1 (B) × W 1,∞ (0, T
).
Since v0 ∈ H 2 (B) and f ∈ Lp ((0, T ) × B), we get

v ∈ W 1,p (0, T) × B ∩ Lp 0, T : W 2,p (B)
(see, e.g., L ADYZENSKAJA , S OLONNIKOV and U RALSEVA [1991], Theorem 9.1,
] × B), hence
Chapter IV). Since p > 4, W 1,p ((0, T ) × B) ⊂ L∞ ([0, T

u∈W 1,q
(0, T ) × B ∩ Lq 0, T : W 2,q (B) ,
for q ∞. Consequently, we get R ∈ W 2,p (0, T ).
]) ⊂ L∞ (B × [0, T
One obtains from the lemma, in view of W0 (B × [0, T
1,p ]) (for
p > 4) the following corollary.
]).
C OROLLARY 5.1. u, v ∈ L∞ (B × [0, T
Utilizing the continuous embedding

W 1,q (0, T ) × B ∩ Lq 0, T : W 2,q (B) ⊂ L2 0, T : W 1,∞ (B) ,

W 1,p (0, T) × B ∩ Lp 0, T : W 2,p (B) ⊂ L2 0, T : W 1,∞ (B) ,
and undoing the change of variables and unknown (4.3)–(4.5) and (5.17), we obtain
C OROLLARY 5.2. Under the assumptions of Theorem 4.1, we have

T

σ 2W 1,∞ (R(t )) + β2W 1,∞ (R(t )) dt k0
0
for some k0 < ∞.
The uniqueness of solutions is established in the next theorem.
T HEOREM 5.1. Let f ∈ Lp (ω0 × (0, T )) with p > 4, and (σ0 − σ , β0 − β) ∈

W 2,s (B(R0 )) ∩ H01 (B(R0 )), for s > 4. Then, there exists a unique solution to (5.1)–
(5.5).
P ROOF. In arguing by contradiction, we assume that there exist two different solutions
(σ1 , β1 , R1 ) and (σ2 , β2 , R2 ). Let

R(t) = min R1 (t), R2 (t) , σ = σ1 − σ2 , β = β1 − β2 .
Then (σ, β, R) satisfies the problem,
∂σ
− dσ + r1 σ + λβ = 0, |x| < R(t), t ∈ (0, T ), (5.15)
∂t
∂β
− dβ + r2 β = 0, |x| < R(t), t ∈ (0, T ), (5.16)
∂t
σ (x, 0) = 0, β(x, 0) = 0, |x| < R0 , (5.17)
σ (x, t) = σ1 (x, t) − σ2 (x, t), |x| = R(t), t ∈ (0, T ), (5.18)
β(x, t) = β1 (x, t) − β2 (x, t), |x| = R(t), t ∈ (0, T ). (5.19)
We introduce a new unknown defined by
z = k1 σ − k2 β,
with
λ
k1 = 1, k2 = if r1 = r2 ,
r1 − r2
1 λ
k1 = , k2 = if r1 = r2 = 0.
2 r1 − 2r2
By construction of z, we have

 ∂z

 − dz + r1 z = 0, |x| < R(t), t ∈ (0, T ),
∂t
z(x, 0) = 0, |x| < R0 , (5.20)



z = k1 σ − k2 β, |x| = R(t), t ∈ (0, T ).
We need the following preliminary result.
L EMMA 5.2. Let z be the solution to the problem (5.20) and β the solution to (5.16),
(5.19), then er1 t z and er2 t β take their maximum and minimum on |x| = R(t).
P ROOF. Multiplying Eq. (5.20) by er1 t , we obtain that er1 t z satisfies


 ∂ r1 t

 e z − d er1 t z = 0, |x| < R(t), t ∈ (0, T ),
∂t
0) = 0, |x| < R0 , (5.21)

 z(x,
 r1 t
e z = er1 t (k1 σ − k2 β), |x| = R(t), t ∈ (0, T ).
In the same way, er2 t β satisfies

 ∂ r2 t

 e β − d er2 t β = 0, |x| < R(t), t ∈ (0, T ),
∂t
β(x, 0) = 0, |x| < R0 , (5.22)


 r2 t
e β = e (β1 − β2 ),
r 2 t |x| = R(t), t ∈ (0, T ).
Applying Corollary 5.1, we obtain that er1 t z and er2 t β are bounded. Let

z∗∗ = max er1 t z(x, t), t ∈ [0, T ], x ∈ ∂B R(t) ,

z∗∗ = min er1 t z(x, t), t ∈ [0, T ], x ∈ ∂B R(t) ,

β ∗∗ = max er2 t β(x, t), t ∈ [0, T ], x ∈ ∂B R(t) ,

β∗∗ = min er2 t β(x, t), t ∈ [0, T ], x ∈ ∂B R(t) .
Notice that z∗∗ 0, β ∗∗ 0, z∗∗ 0 and β∗∗ 0. Let Tk and T k be defined by

s, if s > k, k, if s k,
Tk (s) = and T k (s) =
k, if s k, s, if s < k.
Taking T0 (er1 t z − z∗∗ ) as test function in (5.21) and integrating by parts over B(R(t)),
we arrive after some manipulations at

d rt 2
T0 e 1 z − z∗∗ dx 0.
dt B(R(t ))
We deduce that er1 t z takes his maximum on |x| = R(t). In the same way, taking
T 0 (er1 t z − z∗∗ ) as test function, we obtain
z∗∗ er1 t z z∗∗ . (5.23)

The proof of
β∗∗ er2 t β β ∗∗ , (5.24)

is analogous.
E ND OF THE PROOF OF T HEOREM 5.1. Given t ∈ [0, T ], we can assume, without lost
of generality, that R1 (t) R2 (t). Consider

R1 (t)Ṙ1 (t) − R2 (t)Ṙ2 (t) =
2 2
S(σ1 , β1 ) − S(σ2 , β2 ) dx
B(R(t ))

− S(σ2 , β2 ) dx.
R1 (t )<|x|<R2 (t )
Since S is bounded, then

S(σ2 , β2 ) dx N R13 (t) − R23 (t) M R1 (t) − R2 (t),

R1 (t )<|x|<R2 (t )
where M depends only of |S|L∞ . Since S is Lipschitz continuous, integrating in time,

it results
T

S(σ1 , β1 ) − S(σ2 , β2 ) dx dt
0 B(R(t ))
T
|S|W 1,∞ (R2 ) sup |σ | + sup |β| dx dt
0 B(R(t ))
T
1
k0 sup |z + k2 β| + sup |β| dx dt
0 B(R(t )) k1

T
C sup |z| + sup |β| dx dt
0 B(R(t ))

T
C supe−r1 t er1 t z + supe−r2 t er2 t β dx dt
0 B(R(t ))

T
C e|r1 |T super1 t z + e|r2 |T super2 t β dx dt
0 B(R(t ))

T
k3 super1 t z + super2 t β dx dt.
0 B(R(t ))
From Lemma 5.2, we know

T
3π T 3
super1 t z(x, t) dx dt er1 T R (t) sup z(x, t) dt.
0 B(R(t )) 4 0 |x|=R(t )
By Corollary 5.2, we deduce that

T

σ2 2W 1,∞ (B(R(t ))) + β2 2W 1,∞ (B(R(t ))) dt K0 ,
0
and consequently,
T
z2W 1,∞ (B(R(t ))) dt K.
0
Since

er1 t z(x, t) = er1 t k1 σ2 (x, t) − σ − k2 β2 (x, t) − β , on |x| = R(t),
we deduce

3π T
er 1 T R 3 (t) sup z(x, t) dt
4 0 |x|=R(t )
T
k4 σ2 W 1,∞ (B(R2 (t ))) + β2 W 1,∞ (B(R2(t )))R1 (t) − R2 (t) dt
0

T
k4 sup R1 (t) − R2 (t)T 1/2 σ2 2W 1,∞ (B(R
2 (t )))
0<t <T 0

+ σ2 2W 1,∞ (B(R dt
2 (t )))

k sup R1 (t) − R2 (t)T 1/2.
0<t <T
In the same way,
T

k3 sup |β| dx dt k sup R1 (t) − R2 (t)T 1/2 .
0 B(R(t )) 0<t <T
Then
T
R 2 (t)Ṙ1 (t) − R 2 (t)Ṙ2 (t) dt C0 sup R1 (t) − R2 (t) T + T 1/2 . (5.25)
1 2
0 0<t <T
Let δ = maxt ∈[0,T ] {R1 (t) − R2 (t)} then
3
R (t) − R 3 (t) 3C0 δ T + T 1/2 ,
1 2
since |R13 (t)−R23 (t)| 3R02 |R1 (t)−R2 (t)|, it follows δ k0 δ(T +T 1/2 ). Furthermore,
if T < T1 = min{1/4k02, 1}, necessarily R1 (t) = R2 (t). Since er1 t z and er2 t β take their
maximum and minimum on R(t) = R1 (t) = R2 (t), and R(t) is zero, β = 0 and z = 0,
and we deduce σ = 0. Repeating the process, starting now from T1 , we conclude the
uniqueness of solutions for any T > 0 provided R(T ) > 0.
R EMARK 5.1. Other qualitative properties of the solutions of this type of models have
been studied in the literature by different authors. In particular, we mention the study
of the asymptotic behavior, when t → +∞ (see, e.g., B YRNE and C HAPLAIN [1996a],
F RIEDMAN and R EITICH [1999], C UI and F RIEDMAN [2000], C UI and F RIEDMAN
[2001]) and the continuous dependence and bifurcation phenomena with respect to pa-
rameters (see, e.g., B YRNE and C HAPLAIN [1995], F RIEDMAN and R EITICH [2001],
F RIEDMAN , H U and V ELÁZQUEZ [2001], among others).
5.2. Uniqueness of solutions with radial symmetry
Let (σ̂ , β̂) be a solution of problem (3.1)–(3.5) without forcing term (i.e., f = 0). We
assume the solution is radially symmetric and define σ = σ̂ − σ , β = β̂ − β and r = |x|.
Then (σ, β) verifies

 ∂σ d1 ∂ ∂

 − 2 r 2 σ ∈ g1 (σ, β), 0 < r < R(t), 0 < t < T ,

 ∂t r ∂r ∂r





 ∂β d2 ∂ 2 ∂ β = g (σ, β), 0 < r < R(t), 0 < t < T ,

 − r 2

 ∂t r 2 ∂r ∂r

 R(t )



 2 dR(t) =
 R(t) S(σ, β)r 2 dr, 0 < t < T ,
dt 0 (5.26)

 ∂σ ∂β

 (0, t) = 0, (0, t) = 0, 0 < t < T,



 ∂r ∂r



 σ R(t), t = 0, β R(t), t = 0,

 0 < t < T,



 R(0) = R0 ,




σ (r, 0) = σ0 (r), β(r, 0) = β0 (r), 0 < r < R0 ,
where gi are given by

g1 (σ, β) = − (r1 + λ) σ + σ − r1 σB + β + β H σ + σ − σn , (5.27)

g2 (σ, β) = −r2 β + β . (5.28)
We will assume in this subsection that
1,∞ 2
S ∈ Wloc R , (5.29)
S is an increasing function in σ and decreasing in β, (5.30)
r1 σB − β
σn (5.31)
r1 + λ
and the initial data (σ0 = σ̂ − σ , β0 = β̂0 − β ) belong to H 2 (0, R0 ) and satisfy
∂σ0 ∂β
(0, t) = 0, (0, t) = 0, 0 < t < T , (5.32)
∂r
∂r
σ R(t), t = 0, β R(t), t = 0, 0 < t < T . (5.33)
T HEOREM 5.2. There is, at most, one solution to (5.26).
We will use some earlier results in the proof.

L EMMA 5.3. Every solution (σ, β) of the problem (5.26) is bounded and satisfies σn
σ σB and −β β max{β0 } provided σn σ0 σB and −β β0 .
P ROOF. By the “integrations by parts formula” (justifying the multiplication of the

equation by T0 (σ − σB ) and posterior integrations in time and space, see A LT and
L UCKHAUS [1983], Lemma 1.5), we have
t R(s)
1 R(t ) 2
T0 (σ − σB ) r 2 dr g1 (σ, β)T0 (σ − σB )r 2 dr ds.
2 0 0 0
Since

− (r1 + λ) σ + σ − r1 σB + β + β H σ + σ − σn T0 (σ − σB )

= −(r1 + λ)T0 (σ − σB )2 − (r1 + λ) σB + σ − r1 σB + β − β T0 (σ − σB )

− λσB + (r1 + λ)σ + β + β T0 (σ − σB )
1 0 2 2
T 0 β + β T0 (σ − σB ) T β +β + T0 (σ − σB ) ,
2
we obtain
R(t ) t R(s)
0 2
T0 (σ − σB )2 r 2 dr T β + β + T0 (σ − σB )2 r 2 dr ds.
0 0 0
(5.34)
In the same way, we consider T 0 (β + β ), and since
2
r2 β + β H σ + σ − σn T 0 β + β r2 T 0 β + β ,
it follows that
R(t ) t
0 2 2 R(s)
T β + β r dr r2 T 0 β + β r 2 dr ds. (5.35)
0 0 0
Adding (5.34) and (5.35), we obtain thanks to Gronwall’s lemma
σ σB and β −β.
Notice that β −β implies β̂ 0.

Let us consider ε > 0 and take T 0 (σ − σn − ε) as test function in the weak formula-
tion, then

1 R(t ) 0 2
T (σ − σn − ε) r 2 dr 0.
2 0
Now, taking limits as ε → 0, one concludes

1 R(t ) 0 2
T (σ − σn ) r 2 dr 0,
2 0
which proves σ σn .
Knowing σ and R, β is well-defined as the unique solution of the equation

∂β d2 ∂ 2 ∂

− 2 r β = −r2 β + β , 0 < r < R(t), 0 < t < T ,
∂t r ∂r ∂r
∂β
β R(t), t = 0, = 0 on 0 < t < T .
∂r
Since β0 −β, it follows that

∂β d2 ∂ ∂
− 2 r 2 β 0,
∂t r ∂r ∂r
and we obtain by maximum principle that β max{β0 }.
C OROLLARY 5.3. There exists a positive constant M such that R(t) R0 eMt and
R (t) R0 MeMT .
P ROOF. The above result shows (σ (r, t), β(r, t)) ∈ [σn , σB ] × [−β, max{β0 }]. Since S
is a continuous function, it attains its maximum (denoted by 3M) on that set. Thus,
R(t )
dR(t)
R 2 (t) 3Mr 2 dr.
dt 0
Integrating the above equation, we have dR(t)/dt MR(t). Finally, the conclusion
follows by Gronwall’s lemma.
R EMARK 5.2. As in the previous subsection the solution (σ, β) of (5.26) satisfies
T

σ 2W 1,∞ (ε,R(t )) + β2W 1,∞ (ε,R(t )) dt C1
0
for all ε > 0.
P ROOF OF T HEOREM 5.2. We argue by contradiction and assume that (σ1 , β1 , R1 )

and (σ2 , β2 , R2 ) are two solutions of the problem. Let R(t) := min{R1 (t), R2 (t)}, σ :=
σ1 − σ2 and β := β1 − β2 be the solution to


 ∂σ d1 ∂ 2 ∂ σ = g (σ , β ) − g (σ , β ), r < R(t), 0 < t < T ,

 − r 1 1 1 1 2 2

 ∂t r 2 ∂r ∂r





 ∂β d2 ∂ ∂

 − 2 r 2 β = g2 (σ1 , β1 ) − g2 (σ2 , β2 ), r < R(t), 0 < t < T ,

 ∂t r ∂r ∂r


∂σ ∂β
(0, t) = 0, (0, t) = 0, 0 < t < T, (5.36)

 ∂r ∂r




 σ R(t), t = σ1 R(t), t − σ2 R(t), t ,
 0 < t < T,





 β R(t), t = β1 R(t), t − β2 R(t), t , 0 < t < T,



 σ (r, 0) = 0, β(r, 0) = 0, 0 < r < R0 .
Now, we state a technical lemma.
L EMMA 5.4. |β| takes the maximum on the boundary R(t) and σ satisfies
R(t ) 2
2
T0 (σ − σ ∗ ) r 2 dr T C max {β} ,
0 t ∈[0,T ]
where

σ ∗ = max σ R(t), t .
t ∈[0,T ]
P ROOF. Let us consider β∗ = min{0, β(R(t), t)} and

g2 (β1 ) − g2 (β2 ) = −r2 β1 − β − β2 − β = −r2 β,
then

g2 (β1 ) − g2 (β2 ) T 0 (β − β∗ ) = −r2 βT 0 (β − β∗ ) 0.
Multiply the equation by T 0 (β − β∗ ), we get
R(t )
0 2
T (β − β∗ ) r 2 dr 0
0
and obtain β β∗ . In the same way, we prove that β takes its maximum on R(t).
Let us consider
g1 (σ1 , β1 ) − g1 (σ2 , β2 )

= − (r1 + λ) σ1 + σ − r1 σB + β1 + β H σ1 + σ − σn

− (r1 + λ) σ2 + σ − r1 σB + β2 + β H σ2 + σ − σn

= (r1 + λ) σ1 + σ − σn H σ1 + σ − σn − σ2 + σ − σn H σ2 + σ − σn

+ −(r1 + λ)σn + r1 σB − β H σ1 + σ − σn − H σ2 + σ − σn

− β1 H σ1 + σ − σn − β2 H σ2 + σ − σn .
Since (σ + σ − σn )H (σ + σ − σn ) is an increasing function of σ , we obtain that

− σ1 + σ − σn H σ1 + σ − σn − σ2 + σ − σn H σ2 + σ − σn
× T0 (σ1 − σ2 − σ ∗ ) 0.
Since −(r1 + λ)σn + r1 σB − β 0, it follows that

−(r1 + λ)σn + r1 σB − β H σ1 + σ − σn − H σ2 + σ − σn
× T0 (σ1 − σ2 − σ ∗ ) 0.
Then

g1 (σ1 , β1 ) − g1 (σ2 , β2 ) T0 (σ1 − σ2 − σ ∗ )

− β1 H σ1 + σ − σn − β2 H σ2 + σ − σn T0 (σ1 − σ2 − σ ∗ )

−(β1 − β2 )H σ2 + σ − σn T0 (σ1 − σ2 − σ ∗ )
−T 0 (β1 − β2 )T0 (σ1 − σ2 − σ ∗ ) −β∗ T0 (σ1 − σ2 − σ ∗ ).
Multiplying the equation, as before, by T0 (σ − σ ∗ ), we get

R(t ) t R(s) 2
2 ∂
T0 (σ − σ ∗ ) r 2 dr + T0 (σ − σ ∗ ) r 2 dr ds
0 0 0 ∂r
t R(s)

= g1 (σ1 , β1 ) − g1 (σ2 , β2 ) T0 (σ − σ ∗ )r 2 dr ds
0 0
t R(s)
− β∗ T0 (σ − σ ∗ )r 2 dr ds
0 0
t R(s)
TC 2

β∗2 + λ T0 (σ1 − σ2 − σ ∗ ) r 2 dr ds.
λ 0 0
Now, we choose λ such that
R(s)
2
λ T0 (σ1 − σ2 − σ ∗ ) r 2 dr
0
R(s) ∂
2
− T0 (σ − σ ∗ ) r 2 dr 0 a.e. t ∈ (0, T ),
0 ∂r
then,
R(t ) 2
T0 (σ − σ ∗ ) r 2 dr T Cβ∗2
0
holds, which ends the proof.
E ND OF THE PROOF OF T HEOREM 5.2. Let us define

δ = max R1 (t) − R2 (t) 0,
t ∈[0,T ]
and consider
R12 (t)R1 (t) − R22 (t)R2 (t)
R(t )

= S(σ1 , β1 ) − S(σ2 , β2 ) r 2 dr
0
R1 (t ) R2 (t )
+ S(σ1 , β1 )r dr −
2
S(σ2 , β2 )r 2 dr. (5.37)
R(t ) R(t )
By (5.29) and Lemma 5.3, we obtain
Ri (t )

S(σi , βi )r dr Mδ (for i = 1, 2),
2
(5.38)

R(t )
where

M = max S(σ, β) for any (σ, β) ∈ [σn , σB ] × β, max{β0 } .
(5.29) and (5.30) imply
R(t )
2 R(t )
S(σ1 , β1 ) − S(σ2 , β2 ) r dr C T0 (σ ) − T 0 (β) r 2 dr.
0 0
Since T0 (σ ) T0 (σ − σ ∗ ) + σ ∗ and −T 0 (β) −β∗ , we obtain

R(t )

S(σ1 , β1 ) − S(σ2 , β2 ) r 2 dr
0
R(t )
C T0 (σ − σ ∗ ) + σ ∗ − β∗ r 2 dr
0
R(t ) 1/2
∗ 2 2 ∗
C T0 (σ − σ ) r dr + σ − β∗ .
0
By Lemma 5.4, it follows that
R(t ) 1/2
∗ 2 2

C T0 (σ − σ ) r dr + σ − β∗ C σ ∗ − (T + 1)β∗ .
∗
0
Since σi (Ri (t), t) = 0 (for j = 1 or 2), we obtain

σ R(t), t σi W 1,∞ (R(t ),Ri (t )) R1 (t) − R2 (t),
i=1,2

β R(t), t βi W 1,∞ (R(t ),Ri (t )) R1 (t) − R2 (t)
i=1,2
and then
R(t )

S(σ1 , β1 ) − S(σ2 , β2 ) r 2 dr C(T + 2)δ. (5.39)
0
Integrating in time in (5.37), we get thanks to (5.38) and (5.39) that
R13 (t) − R23 (t) T C(T + 2)δ + 2T Mδ. (5.40)
On the other hand, one has

R13 (t) − R23 (t) = R1 (t) − R2 (t) R12 + R1 R2 + R13 .
We can assume without lost of generality that δ = R1 (t0 )−R2 (t0 ) (for some t0 ∈ [0, T ]),
hence
R13 (t) − R23 (t) 4R 2 δ.
Substituting this into (5.40) leads to δ k0 δT . Furthermore, taking T1 < 1/k0 necessi-
tates R1 (t) = R2 (t) for any t ∈ [0, T1 ]. Since |β| takes its maximum at R(t) = R1 (t) =
R2 (t) (and this maximum is 0), we get that β = 0. Substituting in (5.36) and taking σ
as test function, we obtain
R(t ) t R(s)

σ 2 r 2 dr g1 (σ1 , β1 ) − g1 (σ2 , β2 ) σ r 2 dr ds.
0 0 0
As in Lemma 5.4, since (σi + σ i − σn )H (σi + σ − σn ) is a increasing function of σ ,
we obtain by (5.27) and Lemma 5.3 that (g1 (σ1 , β1 ) − g1 (σ2 , β2 ))σ 0, which prove
σ = 0.
Repeating the above process, starting now from T1 , we get the uniqueness of solutions
for arbitrary T > 0, provided R(T ) > 0.
6. Approximate controllability
In this section we study the controllability of distribution of nutrients (in the usual weak
sense of parabolic system) by the internal localized action of inhibitors. The main results
of this section is the following theorem.
p
T HEOREM 6.1. Given T > 0, ω0 ⊂ B(R0 exp{−SL∞ T }), ε > 0, and σ̂ d ∈ Lloc (R3 ),
for some p > 1, there exists f ∈ Lp ((0, T ) × ω0 ) such that, if (σ, β, R) is the solution
of the problem (5.1)–(5.5), then

σ (T ) − σ d p ε, (6.1)
L (B(R(T )))
where σ d := σ̂ d χB(R(T )) .
Due to some technical reasons, we shall prove the theorem firstly for p > 4. This
assumption is a prerequisite in order to obtain the boundedness of the solution in the
proof of Lemma 5.1 in view of the Sobolev compact embedding W 1,p ((0, T ) × B) ⊂
L∞ ((0, T ) × B). Finally, we prove the theorem for any p > 1 by Hölder inequality.
We shall establish the result in several steps. For n ∈ N, we start by assuming Rn (t)
prescribed and look for a control fn in ω0 such that the solution (σn , βn ) of problem
(5.1), (5.2), (5.4) and (5.5), satisfies (6.1). Then we obtain Rn+1 and fn+1 from (σn , βn )
which allows us to find (σn+1 , βn+1 ). The proof of the theorem relies mostly on methods
introduced in the study of approximate controllability (notion attributed to conclusions
such as (6.1)) by different authors (see L IONS [1990], L IONS [1991], FABRE , P UEL
and Z UAZUA [1995], G LOWINSKI and L IONS [1995] and D ÍAZ and R AMOS [1995]).
Iterating the process, we obtain a sequence (Rn , fn , σn , βn ) such as we shall show pos-
sesses a subsequence that converges to the searched control f and the associate solution
of problem (5.1)–(5.5).
The next result shows the conclusion of Theorem 6.1 (the so-called approximate
controllability in Lp ) under some particular assumptions (mainly, R(t) is a priori pre-
scribed).
P ROPOSITION 6.1. Let ω0 ⊂ B(R0 exp{−SL∞ T }) and σ0 = β0 = σ = β = 0. Let

R ∈ W 1,∞ (0, T ) a given function such that R(0) = R0 , |Ṙ| SL∞ R0 exp{|S|L∞ T }.
Then, given σ̂ d ∈ L2loc (R3 ), there exists f ∈ Lp (ω0 × (0, T )), with p > 4, such that, if
(σ, β) is the solution of problem (5.1), (5.2), (5.4) and (5.5), then

σ (T ) − σ d p ε,
L (B(R(T )))
where σ d = σ̂ d |B(R(T )) .

P ROOF. Let p = p/(p −1) and consider the functional J : Lp (B(R(T ))) → R defined
by

1 T
J ϕ0 = ψ(x, t)p dx dt + εϕ 0 p − σ d ϕ 0 dx,
p 0 ω0 L (B(R(T )))
B(R(T ))
p
where ϕ0 ∈ L (B(R(T ))), and (ϕ, ψ) is the solution to the adjoint problem
∂ϕ
− − dϕ + r1 ϕ = 0, |x| < R(t), t ∈ (0, T ), (6.2)
∂t
∂ψ
− − dψ + r2 ψ + λϕ = 0, |x| < R(t), t ∈ (0, T ), (6.3)
∂t
ϕ(x, T ) = ϕ0 (x), ψ(x, T ) = 0, |x| < R(T ), (6.4)
ϕ(x, t) = 0, ψ(x, t) = 0, |x| = R(t), t ∈ (0, T ). (6.5)
We point out that the existence of a weak solution (ϕ, ψ) of (6.2)–(6.5) can be obtained
as in Section 5, by employing (4.3)–(4.5) and (5.6).
In order to prove the uniqueness of solutions by contradiction, we assume that there
exist two solutions (ϕ1 , ψ1 ), (ϕ2 , ψ2 ). Then ϕ := ϕ1 − ϕ2 satisfies (6.2) and taking

|ϕ|p −2 ϕ as test function and integrating by parts it follows that

d p
− |ϕ| dx r1 |ϕ|p dx.
dt B(R(t )) B(R(t ))
We obtain ϕ = ϕ1 − ϕ2 = 0 by Gronwall’s lemma. Having proved ϕ ≡ 0, in the same

way, ψ := ψ1 −ψ2 satisfies (6.3) and taking |ψ|p −2 ψ as test function, we obtain ψ ≡ 0,
which proves the uniqueness.
Let us assume that J is convex, continuous and coercive (in the sense that
lim inf J → ∞ as ϕ 0 Lp (B(R0 )) → ∞), facts, which shall be proved at the end of the
proposition. Then J takes a minimum ϕ0 (see B REZIS [1983], Corollary III.20). More-
over, if (ξ, ζ ) is the solution of the problem (6.2)–(6.5) with datum (ξ 0 , 0), we have
T

|ψ|p −2 ψζ dx dt − σ d ξ 0 dx
0 ω0 B(R(T ))

0 1−p 0 p −2 0 0
+ εϕ Lp (B(R(T ))) ϕ ϕ ξ dx = 0. (6.6)
B(R(T ))
Multiplying (5.1), (5.2) by (ξ, ζ ), integrating by parts and applying Leibnitz theorem,
we arrive at
T T T
∂ξ
− σ, dt − d σ, ξ dt + r1 σ ξ dx dt
0 ∂t 0 0 B(R(t ))
T T T
∂ζ
+ λβξ dx dt − β, dt − d β, ζ dt
0 B(R(t )) 0 ∂t 0
T T
T
+ r2 βζ dx dt − f ζ dx dt + σ ξ dx 0
0 B(R(t )) 0 ω0 B(R(t ))

T
+ βζ dx 0 = 0,
B(R(t ))
1,p
where , is the duality product W0 (B(R(t))) × W −1,p (B(R(t))). We obtain from
the choice of (ξ, ζ ) and σ (0, x) = β(0, x) = 0 that
T
− f ζ dx dt + σ (T )ξ 0 dx = 0. (6.7)
0 ω0 B(R(T ))
Let us take

f := |ψ|p −2 ψ.
Substituting this into (6.7) and using (6.6), one has

1−p
σ (T ) − σ d ξ 0 dx + εϕ 0 Lp (B(R(T ))) |ϕ 0 |p −2 ϕ 0 ξ 0 dx = 0,
B(R(T )) B(R(T ))
p
for all ξ 0 ∈ L (B(R(T ))). Taking
1
ξ 0 = σ (T ) − σ d p −1 ∈ Lp B R(T ) ,
we obtain in view of p = 1 + 1/(p − 1) that

σ (T ) − σ d p p
L (B(R(T )))

0 1−p 0 p −2 0 1
−1
= εϕ ϕ
Lp (B(R(T )))
ϕ |σ (T ) − σ d | p −1 σ (T ) − σ d dx.
B(R(T ))
By Hölder inequality, we have

0 1−p 0 p −2 0 1
−1
ϕ ϕ ϕ |σ (T ) − σ d | p −1 σ (T ) − σ d dx
Lp (B(R(T )))
B(R(T ))
p−1
σ (T ) − σ d p L (B(R(T )))
,
which leads to

σ (T ) − σ d p ε
L (B(R(T )))
and the conclusion holds.
So, it only remains to check the mentioned properties of J :
J is convex. We can write J as the sum of the functionals

J1 ϕ 0 := − σ d ϕ 0 dx, J2 ϕ 0 := εϕ 0 Lp (B(R(T ))) ,
B(R(T ))
T
1
J3 ϕ 0 := |ψ|p dx dt.
p 0 B(R(t ))
First, we shall see that J3 is convex. Let (ϕ1 , ψ1 ) and (ϕ2 , ψ2 ) be the solutions to (6.2)–
(6.5) with datum ϕ10 , ϕ20 ∈ Lp (B(R(T ))), respectively. Then, since the system is linear,
we get, for α ∈ (0, 1),

0 1 T
J3 αϕ1 + (1 − α)ϕ2 =
0 αψ1 + (1 − α)ψ2 p dx dt
p 0 B(R(t ))
and then

J3 αϕ10 + (1 − α)ϕ20 − αJ3 ϕ10 − (1 − α)J3 ϕ20

1 T
= αψ1 + (1 − α)ψ2 p − α|ψ1 |p − (1 − α)|ψ2 |p dx dt.
p 0 B(R(t ))
Since p > 1, we obtain

αψ1 + (1 − α)ψ2 p − α|ψ1 |p − (1 − α)|ψ2 |p 0,
and integrating, we have

1 T
αψ1 + (1 − α)ψ2 p − α|ψ1 |p − (1 − α)|ψ2 |p dx dt 0,

p 0 B(R(t ))
which proves the convexity of J3 . Finally, J1 is linear and so convex and since
·Lp (B(R(T )) is convex, J2 is also convex.
J is continuous. By construction, J1 and J2 are continuous. We are going to prove

that J3 is also continuous. Let ϕn0 ∈ Lp (B(R(T ))) such that ϕn0 → ϕ 0 and let (ϕn , ψn ),
(ϕ, ψ) be the solutions to (6.2)–(6.5) with datum ϕn0 and ϕ 0 . Subtracting both systems
and taking

p |ϕ − ϕn |p −2 (ϕ − ϕn ), p |ψ − ψn |p −2 (ψ − ψn )
as test function, using the integration by parts formula (see, e.g., A LT and L UCKHAUS
[1983]) and Young’s inequality, we arrive at

∂
− |ϕ − ϕn |p + |ψ − ψn |p dx
∂t B(R(t ))

+ r1 p − |λ| |ϕ − ϕn |p dx + r2 p − |λ| |ψ − ψn |p dx 0.
B(R(t )) B(R(t ))
Let Xn be defined by
p p
Xn (t) = ϕ − ϕn + ψ − ψn ,
Lp (B(R(t ))) Lp (B(R(t )))
then,
p
−Xn (t) CXn (t), t ∈ (0, T ), Xn (T ) = ϕn0 − ϕ 0 Lp (B(R(T ))
are satisfied, where C = max{−r1 p + |λ|, −r2 p + |λ|}. Thus, we obtain

0 Xn (t) Xn (T )e−C(t −T ) .
Since

0 |ψ − ψn |p dx Xn (t),
ω0
we conclude by integrating over [0, T ] and taking limits as n → ∞ that

T T

|ψ − ψn |p dx dt Xn (t) dt → 0,
0 ω0 0
which proves the continuity of J3 .

J is coercive. Let ϕn0 ∈ Lp (B(R(T ))) such that ϕn0 Lp (B(R(T ))) → ∞, when n → ∞.
We claim
J (ϕn0 )
lim inf ε.
n→∞ ϕ p
0
n L (B(R(T )))
Let
J (ϕn0 )
I := lim inf −σ d Lp (B(R(T ))) .
n→∞ ϕ 0 p
n L (B(R(T )))
Then, there exists a minimizing subsequence (which we call again by ϕn0 ) such that
J (ϕn0 )
lim = I.
n→∞ ϕn0 Lp (B(R(T )))
We define
ϕn0
ϕ̄n0 := ,
ϕn0 Lp (B(R(T )))
and let (ϕ̄n , ψ̄n ) be the solution to (6.2)–(6.5) with data (ϕ̄n0 , 0). Since the system is
linear, we have
1
(ϕ̄n , ψ̄n ) = (ϕn , ψn ).
ϕn0 Lp
Then

J (ϕn0 ) p −1 T
= ϕn0 |ψ̄n |p dx dt − σ d ϕ̄n0 dx + ε.
ϕn0 Lp (B(R(T ))) 0 ω0 B(R(T ))
Now, it is clear that, if

T
p
lim inf ψ̄n dx α0 , (6.8)
n→∞ 0 ω0
for some positive α0 , then
J (ϕn0 ) p −1
α0 ϕn0 Lp (B(R(T ))) + ε − σ d Lp (B(R(T ))) → ∞
ϕn Lp (B(R(T )))
0
as n → ∞, which proves the property. Let us assume that

T

lim inf |ψ̄n |p dx = 0.
0 ω0
Then there exists a subsequence ψ̄ni such that

T

|ψ̄ni |p dx dt → 0,
0 ω0
p
therefore ψ̄ni → 0 in L (ω0 × [0, T ]). Taking (0, ζ ) as test function in (6.3), where
ζ ∈ Cc2 ((0, T ) × ω0 ), we obtain
T T
∂ζ
ψ̄ni dx dt − ψ̄ni ζ dx dt
0 ω0 ∂t 0 ω0
T T
− r2 ψ̄ni ζ dx dt + λ ϕ̄ni ζ dx dt = 0.
0 ω0 0 ω0
Taking limits, we conclude that

T
ϕ̄ni ζ dx dt → 0, (6.9)
0 ω0
where ϕ̄ni is the solution to


 ∂ ϕ̄ni

− − dϕ̄ni − r1 ϕ̄ni = 0, |x| < R(t), t ∈ (0, T ),
∂t
ϕ̄ni (t, x) = 0, |x| = R(t), t ∈ (0, T ), (6.10)



ϕ̄ni (T , x) = ϕ̄ ,
0 |x| < R0 .
Repeating the change of (5.6) and introducing the unknown

ūni (x̃, t˜) := ϕ̄ni R t (t˜) x̃, t (t˜) ,
we obtain

 ∂ ūni 2
 − ∂ t˜ − dūni − R R x̃ · ∇ ūni + R r1 ūni = 0, B × (0, T ),
 2
 ūni (x̃, t˜) = 0, ∂B × (0, T), (6.11)



ūni (x̃, T) = ū0ni (x̃) = ϕ̄n0i (x̃R0 ), x̃ ∈ B.

Since ū0ni ū0 belongs to Lp (B), it follows that ūni ū (the solution of (6.11) with

ū0 = ϕ̄ 0 ). By (6.9), ūni → 0 weakly in Lp (B( ω0 )), where ω0 is an open subset of
B such that ω0 ⊂ ω0 for all 0 t˜ T. By the unique
ω0 . Consequently, ū ≡ 0 on
continuation of the solution to Eq. (6.11) (see F RIEDMAN [1964], C HI -C HEUNG P OON
[1996], Theorem 1.1 ), we deduce that ū = 0 in B × (0, T ), which implies ū0 ≡ 0 and
ϕ̄ ≡ 0 by uniqueness of (6.11). Furthermore,
0

− σ d ϕ̄ 0 dx = 0
B(R(T ))
and I = ε, which proves the coerciveness of J .
P ROOF OF T HEOREM 6.1. We consider the function θ : C 1 ([0, T ]) → H 2 (0, T ),

θ (R ∗ ) = R, where R is defined by

R 2 (t)Ṙ(t) = S σ + σ s , β + β s dx, R ∗ (0) = R0 ,
B(R ∗ (t ))
where (σ s , β s ) is the solution to the problem (5.1), (5.2), (5.4) and (5.5), with f ≡ 0,
and initial data σn−1s
(x, 0) = σ0 (x), β s (x, 0) = β0 (x), and (σ, β) is the solution men-
tioned in Proposition 6.1. Since S is bounded, R ∈ W 1,∞ (0, T ). By Proposition 6.1, for
each R ∗ there exists a minimum function ϕn0 which minimize the functional

1 T
J ϕ 0 := |ψ|p dx dt + εϕ 0 Lp (B(R ∗(T ))) − σ d ϕ 0 dx,
p 0 ω0 B(R ∗ (T ))
where σ d = σ̂ d χB(R ∗ (T )) . We are going to show that ϕ 0 Lp (B(R ∗ (T ))) is uniformly
bounded. To the contrary, we assume that there exists a sequence ϕn0 such that
ϕn0 Lp (B(R ∗(T ))) → ∞ and get
T
J (ϕn0 ) 1
ϕ 0 p −1
= |ψ̄n |p dx dt
ϕn0 Lp p n Lp (B(R ∗ (T ))) 0 ω0

+ε− σnd ϕ̄n0 dx 0 (6.12)
B(R ∗ (T ))
in view of Jn (ϕn0 ) 0. Since

σnd ϕn0 d d
dx σ p σ̂ p ,
n ∗ L (B(R0 exp{MT }))
B(R ∗ (T )) ϕn Lp (B(R ∗ (T )))
0 L (B(R (T )))
it follows, by (6.12) that

T

|ψ̄n |p dx dt → 0 when n → ∞.
0 ω0
Using the same argument as in the proof of coerciveness of J , we obtain

ϕ̄n0 0 in Lp B R ∗ (T )
and
Jn (ϕn0 )
lim inf ε,
n→∞ ϕn0
which contradicts (6.12). Consequently ϕn0 Lp (B(R ∗(T ))) is uniformly bounded, hence
ϕn Lp (B(R ∗(T ))) is uniformly bounded. Furthermore, the set of controls is uniformly
bounded. Performing the change of (4.3)–(4.5) and (5.6), applying Lemma 5.1, we ob-
tain that θ is continuous and compact. Then, there exists a fixed point (σ, β, R) which
satisfies (5.1)–(5.5) and condition (6.1). Thus the theorem is proved in the case p > 4.
In the case p 4, we consider the control f for any s > 4, for instance f ∈
L5 ((0, T ) × Ω), then
5
3π p(5−p)
σ (T ) − σ d p meas B R(T ) σ (T ) − σ d 5
L (B(R(T ))) 4 L (B(R(T )))
5
3π p(5−p)
ε exp T SL∞ ,
4
setting
− p(5−p)
3π 5
ε=ε exp T SL∞ ,
4
we obtain the theorem.
R EMARK 6.1. Notice that the final observation is made regarding the density σ (T , ·)
and that once we have chosen the control to obtain (6.1). The free boundary, R(t), and
the inhibitor density β(T , ·) are univocally determined.
R EMARK 6.2. There exists a long literature on the application of Optimization and
Control Theory to different mathematical tumor growth models. We refer the interested
reader to the works by S WAM [1984], F ISTER , L ENHART and M C NALLY [1998], B EL -
LOMO and P REZIOSI [2000] and the references therein.
7. Numerical analysis
In this section we establish a numerical solution to the problem (5.1)–(5.5) by employ-

ing a time discretization scheme which is implicit with respect to u and v and explicit
for the free boundary R. We assume radial symmetry, no forcing terms (i.e., f = 0),
and a nonnecrotic core. Let x := r1 /R(t) and

u(x, t) = σ xR(t), t − σ , v(x, t) = β xR(t), t − β.
Then, problem (3.1)–(3.5) becomes

∂u d1 ∂ ∂ R ∂u
= 2 2 x2 u + x − r1 u − λv + r1 σ + λβ, (0, 1) × (0, T ),
∂t x R ∂x ∂x R ∂x

∂v d2 ∂ 2 ∂v R ∂
= 2 2 x +x v − r2 v + r2 β, (0, 1) × (0, T ),
∂t x R ∂x ∂x R ∂x
t 1
R(t) = R0 exp x 2 S(u, v) dx dt , t > 0,
0 0
ux (0, t) = vx (0, t) = u(1, t) = v(1, t) = 0, t > 0,
R(0) = R0 , u(x, 0) = u0 (x), v(x, 0) = v0 (x), x ∈ (0, 1).
7.1. Time discretization
Let N ∈ N, n = 1, . . . , N and tn = n(T /N). We introduce the approximations
un (x) ≈ u(x, tn ), v n (x) ≈ v(x, tn ), R n ≈ R(tn ),

dR(t)
Ṙ n ≈ in t = tn ,
dt
defined by the following algorithm:
Step 0:

(0.1) R 0 , u0 , v 0 = (R0 , u0 , v0 ),
1 1 2 0 0
(0.2) R 1/2 = R0 + R0 et 0 x S(u ,v ) dx ,
2
1 1 2 0 0
(0.3) Ṙ 0 = R0 x 2 S(u0 , v 0 ) dxR0et 0 x S(u ,v ) dx .
0
Now, for 1 < n N , assuming (R n−1 , un−1 , v n−1 ) be given, we calculate (R n , un ,

vn )
as follows:
Step n:
(n.1)


 v n − v n−1 d2 −2 ∂ 2 ∂ v n + x Ṙ
n−1 ∂

 = x x v n−1

 t n−1
(R ) 2 ∂x ∂x R n−1 ∂x

 − r2 v n + r2 β, in 0 < x < 1,

 n


∂v
(0) = v n (1) = 0,
∂x
(for n = 1, we use R 1/2 ).
(n.2)


 un − un−1 d1 −2 ∂ ∂ n Ṙ n−1 ∂ n−1

 = x x 2 u + x u

 t (R n−1 )2 ∂x ∂x R n−1 ∂x

 − r1 un − λv n + r1 σ + λβ, in 0 < x < 1,

 n


∂u
(0) = un (1) = 0.
∂x
(n.3) We compute R n by integrating according the compound trapezium rule
! n−1 "
1 1 j +1 j +1
R = R0 exp t
n
x 2
S u ,v + S u ,v
j j
dx
2
j =0 0
! # $ "
1
2 1
0 0 n n n−1
j j
= R0 exp t x S u ,v + S u ,v + S u ,v dx .
0 2
j =1
(n.4)
! n−1
1 1 1 j j
Ṙ = R0
n 2
x S u ,v n n
dx exp t x2
S u , v dx
0 2
j =0 0
"
j +1 j +1
+ S u ,v dx
! #
1 1 1 0 0
= R0 2
x S u ,v n n
dx exp T x 2
S u ,v
0 0 2
$ "

n−1
j j
+ S u ,v n n
+ S u ,v dx .
j =1
7.2. Full discretization
We approximate H 1 (0, 1) by space Vh defined by

Vh := φ ∈ C 0 [0, 1] : φ|(xj−1 ,xj ) ∈ P1 , for j = 1, s + 1 ,
where xj = j/(s + 1) and P1 is the space of those polynomials of degree 0 or 1. We
approximate the above implicit–explicit scheme by the system

unh − un−1 D1 ∂ 2 ∂ n Ṙ n−1 ∂ n
h
= x u + x u − r1 unh − λvhn + r1 σ + λβ,
T (xR n−1 )2 ∂x ∂x h R n−1 ∂x h
in 0 < x < 1, n = 1, . . . , N,

vh − vhn−1
n
D2 ∂ 2 ∂ n Ṙ n−1 ∂ n
= x vh + x v − r2 vhn + r2 β,
T (xR n−1 )2 ∂x ∂x R n−1 ∂x h
in 0 < x < 1, i = 1, . . . , N,
∂unh ∂vhn
unh (1) = vhn (1) = 0, = = 0, on x = 0,
∂x ∂x
R(0) = R0 , u0h (x) = uh,0 (x), vh0 (x) = vh,0 (x),
! 1 # $ "
2 1
0 0 n n n−1
j j
Rh = R0 exp T
n
x S uh , vh + S uh , vh + S uh , vh dx .
0 2
j =1
7.3. Weak formulation of the discrete problem
Setting
1
b(ζ, ϕ) = x 2 ζ ϕ dx,
0
the weak formulation of the discrete problem is given by (∀ϕ ∈ Vh )
d1 T T Ṙ n−1 n
(1 + T r1 )b unh , ϕ + n−1 2 b uih x , ϕx − n−1
b x uh x , ϕ
(R ) R
n−1 n−1
= b uh − vh + Γ1 σ + β, ϕ = b un + T −λvhn + r1 σ + λβ , ϕ ,
n
d1 T T Ṙ n−1 n
(1 + T r2 )b vhn , ϕ + n−1 2 b vhn x , ϕx − b x vh x , ϕ
(R ) R n−1

= b vhn−1 + T r2 β, ϕ .
7.4. Numerical experiments
We consider the special case of S(σ, β) = σ − σ̂ , T = 3, N = 501, (i.e., T = 3/500)

and s = 20 (i.e., h = 1/20) with the following choice of the parameters: R0 = 5, D1 =
D2 = 1, Γ1 = Γ2 = σ = β = 1. These values of the parameters have been taken merely
with academical purpose. For other choices see, for instance, B YRNE and C HAPLAIN
[1996a]. In Figs. 7.1, 7.5 and 7.9, we display the computed evolution of the radius of
the tumor for experiments 1 (σ̂ = 0.75), 2 (σ̂ = 1) and 3 (σ̂ = 1.5). In Figs. 7.2, 7.6
and 7.10 we display visualized the computed evolution of the radius of the tumor in two
F IG . 7.1. F IG . 7.2.
F IG . 7.3. F IG . 7.4.
F IG . 7.5. F IG . 7.6.
F IG . 7.7. F IG . 7.8.
dimensions. Figs. 7.3, 7.7 and 7.11 show the computed evolution of the concentration
of nutrients σ . Finally, in Figs. 7.4, 7.8 and 7.12 we exhibit the computed concentration
of the inhibitors β. Numerical simulation of the model (when S = σ − σ̃ ) show us
the importance of the parameter σ̃ in the behavior of the boundary. As it is expected,
a smaller σ̃ produces a faster growth of the boundary. We can see in Figs. 7.1, 7.5 and 7.9
an initial concave growth of the radius that becomes convex after a time (which depends
on σ̃ ). Among other different aspects it can be appreciated that the free boundary is not
necessarily increasing in time (see Fig. 7.1).
F IG . 7.9. F IG . 7.10.
F IG . 7.11. F IG . 7.12.
Acknowledgement
The work of first author was partially supported by the DGES (Spain) project
REN2000/0766 and RTN HPRN-CT-2002-00274.
References
A DAM , J.A. (1986). A simplified mathematical model of tumor growth. Math. Biosci. 81, 229–244.
A DAM , J.A., B ELLOMO , N. (1997). A Survey of Models for Tumor–Immune System Dynamics (Birkhäuser,
Boston).
A LT , H.W., L UCKHAUS , S. (1983). Quasi-linear elliptic–parabolic differential equations. Math. Z. 183, 311–
341.
A NDERSON , A.R.A., C HAPLAIN , M.A.J. (1998). Continuous and discrete mathematical models of tumor-
induced angiogenesis. Bull. Math. Biology 60, 857–899.
ATTALLAH , A.M. (1976). Regulation of cell growth in vitro and in vivo: point/counterpoint. In: Houck, J.C.
(ed.), Chalones (North-Holland, Amsterdam), pp. 141–172.
BAZALIY , B.V., F RIEDMAN , A. (2003). A free boundary problem for an elliptic–parabolic system: applica-
tion to a model of tumor growth. Comm. Partial Differential Equations 28, 517–560.
B ELLOMO , N., P REZIOSI , L. (2000). Modelling and mathematical problems related to tumor evolution and
its interaction with the immune system. Math. Comput. Modelling 32, 413–452.
B ERTUZZI , A., FASANO , A., G ANDOLFI , A., M ARANGI , D. (2002). Cell kinetics in tumour cords studied
by a model with variable cell cycle length. Math. Biosci. 177–178, 103–125.
B REZIS , H. (1983). Analyse Fonctionnelle (Masson, Paris).
B RITTON , N.F., C HAPLAIN , M.A.J. (1993). A qualitative analysis of some models of tissue growth. Math.
Biosci. 113, 77–89.
B YRNE , H.M. (1997a). The effect of time delays on the dynamics of avascular tumor growth. Math.
Biosci. 144, 83–117.
B YRNE , H.M. (1997b). The importance of intercellular adhesion in the development of carciomas. IMA J.
Math. Appl. Med. Biol. 14, 305–323.
B YRNE , H.M. (1999a). A comparison of the roles of localised and nonlocalised factors in solid tumour
growth. Math. Models Methods Appl. Sci. 9, 541–568.
B YRNE , H.M. (1999b). A weakly nonlinear analysis of a model of avascular solid tumour growth. J. Math.
Biol. 39, 59–89.
B YRNE , H.M., C HAPLAIN , M.A.J. (1995). Growth of nonnecrotic tumors in the presence and absence of
inhibitors. Math. Biosci. 130, 151–181.
B YRNE , H.M., C HAPLAIN , M.A.J. (1996a). Growth of necrotic tumors in the presence and absence of
inhibitors. Math. Biosci. 135, 187–216.
B YRNE , H.M., C HAPLAIN , M.A.J. (1996b). Modelling the role of cell–cell adhesion in the growth and
development of carciomas. Math. Comput. Modelling 12, 1–17.
B YRNE, H.M., M ATTHEWS , P. (2002). Asymmetric growth of avascular solid tumors: exploiting symmetries,
in press.
C HAPLAIN , M.A.J. (1996). Avascular growth, angiogenesis and vascular growth in solid tumours: the math-
ematical modelling of the stages of tumour development. Math. Comput. Modelling 23 (6), 47–87.
C HAPLAIN , M.A.J. (1999). Mathematical models for the growth, development and treatment of tumours.
Math. Models Methods Appl. Sci. 9 (4), 171–206 (special issue).
C HAPLAIN , M.A.J., A NDERSON , A.R.A. (1997). Mathematical modelling, simulation and prediction of
tumour-induced Angiogenesis. Invas. Metast. 16, 222–234.
C HAPLAIN, M.A.J., P REZIOSI , P. (2002). Macroscopic modelling of the growth and development of tumour
masses, in press.
228
References 229
C HI -C HEUNG P OON (1996). Unique continuation for parabolic equations. Comm. Partial Differential Equa-
tions 21, 521–539.
C UI , S., F RIEDMAN , A. (1999). Analysis of a mathematical of protocell. Math. Anal. Appl. 236, 171–206.
C UI , S., F RIEDMAN , A. (2000). Analysis of a mathematical model of effect of inhibitors on the growth of
tumors. Math. Biosci. 164, 103–137.
C UI , S., F RIEDMAN , A. (2001). Analysis of a mathematical model of the growth of the necrotic tumors.
Math. Anal. Appl. 255, 636–677.
DE A NGELIS , E., P REZIOSI , L. (2000). Advection–diffusion models for solid tumor evolution in vivo and
related free boundary problem. Math. Models Methods Appl. Sci. 10, 379–407.
D ÍAZ , J.I., R AMOS , A.M. (1995). Positive and negative approximate controllability results for semilinear
parabolic equations. Rev. Real Acad. Cienc. Exact., Fís. Natur. Madrid LXXXIX, 11–30.
D ÍAZ , J.I., T ELLO , L. (1999). A nonlinear parabolic problem on a Riemannian manifold without boundary
arising in Climatology. Collect. Math. 50, 19–51.
D ÍAZ , J.I., T ELLO , J.I. (2003). On the mathematical controllability in a mathematical in a simple growth
tumors model by the internal localized action of inhibitors. Nonlinear Anal.: Real World Appl. 4, 109–125.
D ÍAZ, J.I., T ELLO, J.I. (2004). Mathematical analysis of a simple model for the growth of necrotic tumors in
presence of inhibitors. Int. J. Pure Appl. Math., in press.
FABRE , C., P UEL , J.P., Z UAZUA , E. (1995). Approximate controllability of the semilinear heat equation.
Proc. Roy. Soc. Edinburgh Sect. A 125, 31–61.
F ISTER , K.R., L ENHART , S., M C N ALLY , J.S. (1998). Optimizing chemotherapy in HIV model. Electron.
J. Differential Equations 32, 1–12.
F ONTELOS , M.A., F RIEDMAN , A., H U , B. (2002). Mathematical analysis of a model for the initiation of
angiogenesis. SIAM J. Math. Anal. 33 (6), 1330–1355.
F RIEDMAN , A. (1964). Partial Differential Equations of Parabolic Type (Prentice Hall, New York).
F RIEDMAN , A. (2002). A hierarchy of cancer models and their mathematical challenges. In: Lecture at the
Workshop on Mathematical Models in Cancer, Vanderbilt University, May 3–5.
F RIEDMAN , A., H U , B., V ELÁZQUEZ , J.J.L. (2001). A Stefan problem for a protocell model with symmetry-
breaking bifurcations of analytic solutions. Interfaces and Free Boundaries 3, 143–199.
F RIEDMAN , A., R EITICH , F. (1999). Analysis of a mathematical model for the growth of tumors. J. Math.
Biol. 38, 262–284.
F RIEDMAN , A., R EITICH , F. (2001). Symmetry-breaking bifurcation of analytic solutions to free boundary
problems: An applications to a model of tumor growth. Trans. Amer. Math. Soc. 353, 1587–1634.
F RIEDMAN , A., T ELLO , J.I. (2002). Stability of solutions of chemotaxis equations in reinforced random
walks. Math. Anal. Appl. 272, 138–163.
G LOWINSKI , R., L IONS , J.L. (1995). Exact and approximate controllability for distributed parameter sys-
tems, Part II. Acta Numer., 157–333.
G REENSPAN , H.P. (1972). Models for the growth of solid tumor by diffusion. Stud. Appl. Math. 52, 317–340.
G REENSPAN , H.P. (1976). On the growth and stability of cell cultures and solid tumors. J. Theoret. Biol. 56,
229–242.
H OLMES , M.J., S LEEMAN , B.D. (2000). A mathematical model of tumour angiogenesis incorporating cel-
lular traction and viscoelastic effects. J. Theoret. Biol. 202, 95–112.
L ADYZENSKAJA , O.H., S OLONNIKOV , V.A., U RALSEVA , N.N. (1991). Linear and Quasi-linear Equations
of Parabolic Type, Transl. Math. Monogr. 23 (Amer. Math. Soc., Providence, RI).
L EVINE , H.A., S LEEMAN , B.P. (1997). A system of reaction diffusion equations arising in the theory of
reinforced random walks. SIAM J. Appl. Math. 57, 683–730.
L EVINE, H.A., PAMUK, S., S LEEMAN, B.P., N ILSEN -H AMILTON, M. Mathematical modeling of capillary
formation and development in tumor angiogenesis: Penetration into the stroma, in press.
L EVINE , H.A., S LEEMAN , B.P., N ILSEN -H AMILTON , M. (2000). A mathematical modeling for the roles of
pericytes and macrophages in the initiation of angiogenesis I. The role of protease inhibitors in preventing
angiogenesis. Math. Biosci. 168, 75–115.
L IONS , J.L. (1990). Remarques sur la contrôllabilité approchée. In: Actas de Jornadas Hispano–Francesas
sobre Control de Sistemas Distribuidos, Universidad de Malaga, pp. 77–88.
L IONS , J.L. (1991). Exact Controllability for distributed systems: Some trends ans some problems. Appl.
Indust. Math., 59–84.
O RME , M.E., C HAPLAIN , M.A.J. (1995). Travelling waves arising in mathematical models of tumour an-
giogenesis and tumour invasion. Forma 10, 147–170.
O THMER , H.G., S TEVENS , A. (1997). Aggregation, blowup, and collapse: The ABC’s of taxis in reinforced
random walks. SIAM J. Appl. Math. 57, 1044–1081.
O WEN , M.R., S HERRATT , J.A. (1999). Mathematical modeling of macrophage dynamics in tumors. Math.
Models Methods Appl. Sci. 9, 513–539.
S IMON , J. (1987). Compact sets in the space Lp ((0, T ), B). Ann. Mat. Pura Appl. CXLVI, 65–96.
S LEEMAN , B.D. (1996). Solid tumor growth: A case study in mathematical biology. In: Aston, Ph.J. (ed.),
Nonlinear Mathematics and its Applications (Cambridge Univ. Press, Cambridge), pp. 237–254.
S HOWALTER , R.E. (1996). Monotone Operator in Banach Space and Nonlinear Equations (Amer. Math.
Soc., Philadelphia).
S HYMKO , R.M., G LASS , L. (1976). Cellular and geometric control of tissue growth and mitotic instability.
J. Theoret. Biol. 63, 355–374.
S WAM , G. (1984). Applications of Optimal Control Theory in Biomedicine (Dekker, New York).
T HOMPSON , K.E., B YRNE , H.M. (1999). Modelling the internalisation of labelled cells in tumour spheroids.
Bull. Math. Biol. 61, 601–623.
VALENCIANO , J., C HAPLAIN , M.A.J. (2003a). Computing highly accurate solutions of a tumour angiogen-
esis model. Math. Models Methods Appl. Sci. 13, 747–766.
VALENCIANO, J., C HAPLAIN, M.A.J. (2003b). An explicit subparametric spectral element method of lines
applied to a tumour angiogenesis system of partial differential equations. Math. Models Methods Appl.
Sci., in press.
V RABIE , I.I. (1995). Compactness Methods for Nonlinear Evolutions, second ed. (Longman, Essex).
WARD , J.P., K ING , J.R. (1998). Mathematical modelling of avascular tumor growth II: Modeling growth
saturation. IMA J. Math. Appl. Med. Biol. 15, 1–42.
Human Models for Crash
and Impact Simulation
Eberhard Haug
ESI Software S.A., 99, rue des Solets, BP 80112,
94513 Rungis Cedex, France
E-mail: eha@esi-group.com
URL: http://www.esi-group.com
Hyung-Yun Choi
Hong-Ik University, Seoul, South Korea
E-mail: hychoi@hongik.ac.kr
Stéphane Robin
LAB PSA-Renault, Paris, France
E-mail: stephane.robin@mpsa.com
Muriel Beaugonin
ESI Software S.A., Paris, France
E-mail: mbe@esi-group.com
231
Contents
P REFACE 235
C HAPTER I. I NTRODUCTION 237
1. On the interest, need and difficulties of using human models in virtual passenger car crash tests 237
2. Overview on mechanical dummies and models 253
3. Overview of existing human models for occupant safety 259
C HAPTER II. H UMAN A RTICULATED M ULTI -B ODY M ODELS 271
4. Human Articulated Rigid Body (HARB) models 271

5. The Hill muscle model 275
6. Application of Hill muscle bars 276
7. Application of HARB models 280
C HAPTER III. D EFORMABLE H UMAN M ODELS 287
8. The HUMOS human models for safety 287

9. The fiftieth percentile male H-Model 297
10. The fifth percentile female H-Model 357
A PPENDIX A. BASIC T HEORY OF C RASH C ODES 363
A.1. Overviews on solution methods and finite elements 363

A.2. Explicit solution method outline 366
A.3. Contact treatment outline 370
A PPENDIX B. B IOMATERIAL M ECHANICAL P ROPERTIES 375
B.1. Bone mechanical properties 375

B.2. Ligament mechanical properties 381
B.3. Brain mechanical properties 389
B.4. Joint mechanical properties 389
B.5. Inter-vertebral joint mechanical properties 395
B.6. Articular cartilage and meniscus mechanical properties 399
B.7. Inter-vertebral disc mechanical properties 399
B.8. Muscle mechanical properties 400
B.9. Tendon mechanical properties 403
B.10. Skin mechanical properties 404
B.11. Internal organ mechanical properties 404
233
234 E. Haug et al.
B.12. Quasi-linear visco-elastic solids 410

B.13. Further references on biomaterials 412
A PPENDIX C. T HE H ILL M USCLE M ODEL 413
A PPENDIX D. A IRBAG M ODELS 421
A PPENDIX E. I NTERACTIONS BETWEEN PARTS 427
R EFERENCES 433
Preface
This article deals with the application of computational impact biomechanics to the
consequences of real world passenger car accidents on human occupants, using com-
puter models in numerical simulations with industrial crash codes. The corresponding
developments are illustrated on the subject of safety simulations of human passenger
car occupants. With some adaptations, the developed models apply equally well to the
simulation of pedestrian accidents and to the design for occupant safety of motorbikes,
trucks, railway vehicles, airborne vehicles, seagoing vessels and more.
The human models elaborated in this article belong to the class of finite element
models. They can be adapted, specialized and packaged for other industrial applications
in human ergonomics and comfort analysis and design, in situations where humans
operate at their work place, as military combatants, or in sports and leisure activities
and more. In the medical field, biomechanical human models can serve as a basis for
the simulation and design of orthopedic prostheses, for bone fracture planning, physical
rehabilitation analysis, the simulation of blood flow, artificial blood vessels, artificial
heart valves, bypass operations, and heart muscle activity, virtual organ surgery, etc.
There exists indeed a large overlap, and a pressing urge and opportunity for creating
a synergy of very diverse disciplines, which all deal with the simulation of the biome-
chanical response of the human body.
Most considerations of this article are related to the application of modern crash
codes, which discretize space with the finite element method and which apply the ex-
plicit time integration scheme of the dynamic equations of motion to discrete numerical
models. The reader is assumed to be familiar with the associated basic theory, needed
for the use of such codes.
The article is structured as follows.
Chapter I provides an introduction on the interest, need and difficulties of using hu-
man models in occupant safety design and analysis. It contains a short overview on
mechanical dummies, often used for the design of occupant safety of transport vehicles,
and it summarizes some so far existing biomechanical human computer models.
Chapter II discusses “MB (multi-body)” or “HARB (Human Articulated Rigid
Body)” or “ATB (Articulated Total Body)” models. These simplest human models con-
sist in rigid body segments, joined at the locations of their skeletal articulations, which
can provide gross overall kinematic responses of the human body to static and dynamic
load scenarios. For more detailed investigations, they can serve as a basis for modular
plug-in of more elaborate and deformable segment models, for making zooms on the
detailed response of various body parts. The chapter closes with applications to occu-
235
236 E. Haug et al.
pant safety of HARB models, including the fifth percentile female and a six year old
child model.
Chapter III discusses deformable human models. In a first section, the results of the
first European HUMOS (Human Models for Safety) project (1999–2001) are summa-
rized. The HUMOS-1 project was funded by the European Commission in the Industrial
and Materials Technologies (IMT) program (Brite–EuRam III). In this project the geom-
etry of a near fiftieth percentile human cadaver geometry was acquired in a passenger
car driving posture and human models were derived from the anthropometrical, bioma-
terial and validation database, compiled and generated within this project.
In a second section, a systematic presentation of the generation of human models and
sub-models is given and illustrated on the example of a deformable fiftieth percentile
human model (H-Model). This section first outlines the HARB version of the model
and then the deformable sub-models of the head, skull and brain, the neck and cervical
spine, the torso with the rib cage, thoracic and abdominal organs, the upper extrem-
ity with the shoulder and arm complex, the lower extremity with the knee, thigh and
hip complex and the ankle-foot complex. For each deformable sub-model the relevant
anatomy, the main injury mechanisms, the basic model structure, its calibration and the
basic validations of the models are outlined. A validation of an abdomen model is dis-
cussed in the first section. A final section outlines the emerging deformable models of
the fifth percentile female.
Appendix A gives an overview on the basic theory of explicit solution and on contact
treatment. Appendix B contains data on biomaterials. Appendix C outlines the Hill type
muscle models. Appendix D discusses the numerical entities of air- and bio-bags, used
to simulate protective airbags and hollow organs. Appendix E provides an insight into
the management of the interaction of parts and organs in biomechanical simulation of
the human body.
It is clear that this article can only be an incomplete outline over the fast growing, vast
and stimulating subject of biomechanical (and biomedical) modeling techniques of the
human body. The presented models and methodologies will undoubtedly be upgraded
by the time this article is printed. The interested reader is therefore encouraged to keep
a close watch on the corresponding web sites and the open literature.
C HAPTER I
Introduction
1. On the interest, need and difficulties of using human models in virtual
passenger car crash tests
1.1. Crash design
Crash tests. In car design, standardized “legal laboratory crash tests” are made in or-
der to assess the protective and life saving performance of the car body and its built-in
passive occupant safety devices, such as airbags, protective paddings and seat belts.
Conventionally, the response of car occupants under accidental conditions, as in frontal
crashes, lateral side impacts, rollover accidents, etc., is studied, using re-usable “me-
chanical occupant surrogate devices”, often called “mechanical dummies” or “legal
crash dummies”. According to existing regulations, passenger transport vehicles must
be designed to pass standard crash tests safely to obtain legal certification for selling
them to customers. The achieved safety levels are assessed through the mechanical re-
sponses of the used dummy devices, as recorded by instruments in physical crash tests
carried out in crash laboratories. These recordings are correlated heuristically with hu-
man injury. Safe crash design methodologies have their widest use in passenger car
design, but apply to all road, water and airborne passenger transport vehicles and work-
ing devices. Recent efforts towards “legal virtual testing” try to establish regulatory
frameworks that can be used to replace laboratory tests for the purpose of legal certi-
fication of vehicles with simulation. While desirable for working with dummies, such
process will be mandatory for working with human models since no real world tests can
be made to back up the simulations.
Crash simulation. In numerical passenger car crash simulations, numerical models

of the car structure, the passive safety restraints (seat belts, airbags, cushions) and the
dummy devices are made, the latter as simpler multi-body models, or as more elaborate
deformable finite element dummy models. Care must be taken that the passive safety de-
vices are modeled with enough detail, so that their deployment, deformation and energy
absorption capacities are well represented in the simulation of a car crash. The numer-
ical dummy models are placed inside the numerical models of the vehicle structure,
and their performance under an imposed crash scenario is evaluated. Models of passive
safety devices, such as airbags, seat belts, knee bolsters, etc., will be designed and op-
timized to improve the car safety or crashworthiness performance with respect to the
237
238 E. Haug et al. C HAPTER I
used dummy models. The safety of the car for human occupants is assessed through the
simulated response output of the virtual dummy devices, which are modelled and “in-
strumented” to behave like the real world mechanical dummy devices. If human models
were used instead of models of mechanical occupant surrogates, or dummies, a more
direct access to human injury could be provided.
Crash codes overview. Numerical crash simulations are performed with specialized
crash codes, which were conceived during the eighties of the last century (Pam–Crash,
Radioss, Dyna3D), following an urgent need for economy, safety and speed of passen-
ger car design. This need was expressed by the world’s passenger car manufacturers.
Since the standard safety regulations in all countries became more and more strict, the
conventional methods to hand-make ever lighter new car prototype structures and to
crash test them became increasingly uneconomical, time consuming and unsafe. The
only answer to satisfy the pressure for crashworthiness, safety, quick time to market
and economy of design lay in the emerging methodologies of virtual prototyping and
design, using high performance computing. This is why several commercial crash codes
have emerged, all based in essence on the dynamic explicit finite element method of
structural analysis, which uses the proven finite element method for discretizing space,
and the explicit direct integration scheme of the non-linear equations of motion to dis-
cretize time. One early account of the practical application of a commercial crash code
is given by H AUG and U LRICH [1989].
The numerical models treated by these codes started with the car body-in-white
(mostly steel structures), modeled with thin shells and contacts. Soon increasingly trade
specific models of passive safety devices (airbags, seat belts, knee bolsters, etc.), mod-
eled with cables, bars, joints, membranes, shells and solids followed. Within a few years,
models of mechanical dummies, impact barriers and crash obstacles appeared. Today
numerical models of human occupants are under active development, with worldwide
active support of national agencies for traffic safety. Like always in numerical simula-
tion, a trade-off between computational efficiency, robustness of execution and accuracy
must be found. It is therefore legitimate to create numerical models of the human body
at different levels of discretization, where the less discretized models execute faster to
provide more approximate answers in early design stages, and the more elaborate mod-
els cost more computer time and resources, but provide more information and yield
more accurate results for the final design.
The correct simulation of contact events or collisions is one of the most crucial fea-
tures of crash codes. Collisions can occur between the structure of interest and objects
in its environment, such as contact between a car and a rigid wall, car-to-car contact,
or contact of an occupant with an airbag or seat belt. Contact can also occur between
different parts of a crashed structure, such as between the engine and the car body,
tire-to-wheel case, roof-to-steering wheel, occupant arm to occupant chest. Finally, self-
contact can occur within a single car body component after buckling and wrinkling of its
constituent thin sheet metal parts. The correct and efficient treatment of collision events
is therefore of great importance, and crash codes have been conceived giving great at-
tention to contact algorithms. Early accounts on the conception of such algorithms are
S ECTION 1 Introduction 239
found in H UGHES , TAYLOR , S ACKMAN , C URNIER and K ANOKNUKULCHAI [1976],

H ALLQUIST, G OUDREAU and B ENSON [1985], and others.
1.2. Occupant safety design
Occupant surrogates. In real world crash tests, it is common practice to use mechan-
ical dummies as surrogates for the human vehicle occupants. Mechanical dummies are
instrumented biofidelic occupant surrogate devices, made of metallic, rubber, foam and
plastic materials, that are widely used by car makers in real vehicle crash tests. The
impact of car accidents on human occupants is inferred from the impact performance of
the used mechanical dummies, expressed in terms of standard response measurements,
such as head accelerations, chest deflections, femur loads, etc. These measurements can
be correlated with human injury via so-called injury criteria. The latter give rough in-
sight into the real injuries a human occupant might experience in each studied crash
scenario.
The consistent use of dummies in crash tests is not ideal, because even the best crash
dummies can only approximate the behavior of real humans in a crash. Humans un-
dergo wider trajectories inside a vehicle than dummies. Therefore ever more advanced
dummies are needed to provide more representative injury data. Nevertheless, dummies
and dummy models enabled car manufacturers to very significantly increase the passive
safety performances of their products. Most of the current safety devices were indeed
developed with the well-known Hybrid III frontal dummy, or with the EuroSID 1 side
impact dummy. Since humans cannot be used in real world crash tests, dummies are the
only workable alternative, and their use is mandatory. Crash dummies are under con-
tinuous improvement, and next generation mechanical occupant surrogate devices are
under development (e.g., the THOR dummy developed by NHTSA), often with the help
of numerical simulations using human models.
Human subjects. The direct use of humans in the everyday safety design of transport
vehicles is excluded due to obvious ethical and practical constraints. Some exceptional
uses of live and dead humans (cadavers or PMHS for post mortem human subjects) for
research purposes and for indirect studies of the response of the human body in crash
situations are listed next. All tests involving human volunteers and human cadavers are
subject to very rigorous screening procedures by competent ethics committees in all
countries. Adult persons can dedicate their bodies in case of decease by an act of will
to science. Children can not grasp such an act, and their parents can not, in general,
decide for their children. Child cadavers can therefore hardly be used for destructive
tests. Exceptions may exist in using body scan images of children when the parents give
their consent. Modern non-destructive bone density measurements and similar existing
or emerging techniques can be used to circumnavigate this ethical dilemma.
Human volunteers. Human volunteers can only exceptionally serve in experimental

impact tests. One well known historical contribution were the human tests carried by
US Air Force Colonel John P. Stapp, who studied from 1946 to 1958 the effects of
deceleration on both humans and animals at the Edwards Air Force Base in California
and at the Holloman Air Force Base in New Mexico. Stapp exposed belt restrained
volunteers, including himself, to decelerations of up to 40 g, using rocket driven sleds.
Since 1955 by now yearly Stapp Car Crash Conferences take place (46th by 2002). One
recent example involving human volunteers is given by low energy rear and front impact
crash tests, where the principal effect of neck “whiplash” motions is studied in purely
research oriented projects under medically controlled conditions. O NO [1999] studies
the relationship between localized spine deformation and cervical vertebral motions for
low speed rear impacts using human volunteers. In such exceptional test setups, human
volunteers are subjected to sub-injury rear or front impact equivalent acceleration levels.
In particular, these studies employ X-ray cine-radiography, accelerometer recordings
and electro-myographic recordings on the neck response in very low speed rear-end car
impacts. In such recordings the activation level of the neck muscles can be monitored
via their variable electrical characteristics. The resulting data are used to develop models
to evaluate neck injuries caused by higher-speed rear-end impacts, and to improve the
accuracy of conventional crash-test dummies.
Accidentological studies and accident reconstructions. Such studies can give insight
into crash events after a real life accident has occurred. These investigations can de-
termine what might have happened to the involved human occupants. Accidentological
studies provide data about the ways the accidents occurred, the involved vehicles, vehi-
cle trajectories and collision with obstacles, and data about the injuries and the medical
consequences for the human occupants. Accident reconstruction studies often re-enact
reported crashes in the laboratory, or use numerical simulation of the reported crashes.
In such re-enactions and simulations, mechanical dummies and their models can be em-
ployed. In numerical simulations of the re-enacted crashes, the use of human models is
of particular interest, since the regulations which prescribe the mandatory use of me-
chanical occupant surrogates in certification tests do not apply. Car companies re-enact
reported crashes in order to better understand the causes of injury and to improve the
car design.
Human cadaver tests. Tests with human cadavers (PMHS tests) can be carried out
at the exceptional research level in experimental impact studies. Most cadaver tests
study the basic biomechanical mechanisms that lead to injuries of the human body (e.g.,
S CHMIDT, K ALLIERIS , BARZ , M ATTERN , S CHULZ and S CHÜLER [1978]). In no case
can cadaver tests be used in everyday car design. Only principal injury mechanisms can
be deduced from cadaver tests, and each cadaver tends to be different. Average human
response “corridors” can be derived from test campaigns which may involve many dif-
ferent cadavers, each subjected to the same test. In the past, the design of mechanical
dummies was based largely on the knowledge derived from specific series of different
types of cadaver tests. For example, K ALLIERIS and S CHMIDT [1990] describe the neck
response and injury assessment using cadavers and the US-SID side impact dummy for
far-side lateral impacts of rear seat occupants with inboard-anchored shoulder belts.
Cadaver test results can produce valuable information for the construction of hu-
man numerical models, rather than to be of direct value in everyday car design. For
example, cadaver test studies on the human skull and the mechanisms of brain injury
can clarify the relationship between different types of impacts and the nature and ex-
tent of injury. Tests on the brain and the skull are carried out in order to improve two-
and three-dimensional models of the head for computer simulations, to understand the
mechanisms through which injuries develop in the brain and skull. Neck tests improve
the knowledge about human neck injury tolerance and mechanisms. Pendulum impact
tests on the thorax and pelvis shed light on the response of the skeleton and organs in
frontal and side impacts. Impact tests on the abdomen can give insight in the action of
lap belts on the visceral organs. Upper extremity impact tests yield information about
aggression from side impact airbags. Cadaver test research into leg injuries typically
involves examining intrusion of the fire wall of passenger cars into the occupant com-
partment, the sitting position and kinematics of the occupant, the effectiveness of knee
bolsters, the position of the pedals, and the anatomical nature of these injuries.
Animal tests. Tests which involve life or dead animals are subject to ethics committee
constraints, as are tests involving human cadavers or human volunteers. In order to
discern the different behavior of body segments, organs and biomaterials of the live
organism, animal tests have be performed on live anesthetized animals. Again, such
tests can not serve as a basis for everyday car design, but are sometimes carried out
in purely research type projects where the use of humans is excluded. For example,
some brain injury mechanisms were studied in the past on primates by O NO , K IKUCHI ,
NAKAMURA , KOBAYASHI and NAKAMURA [1980]. Pigs were also used to study the
consequences of chest impacts by K ROELL , A LLEN , WARNER and P ERL [1986].
Humans in crash tests. While humans cannot replace mechanical dummies in real
world crash tests, this is the case in virtual crash simulations. It is therefore of great
potential advantage to build human models, and to use them in accident simulations.
By combining crash analysis and biomechanical analysis, it is possible to advance the
understanding of how injuries occur. This is the most important step towards creating
safer automobiles and safer roads. As a by-product, human models can be used for
improving the design of mechanical occupant surrogate devices.
1.3. Injury and trauma
Humans vs. dummies. Human models represent “bone, soft tissue, flesh and organs”
instead of “steel, rubber, plastic materials and foam”, as it is the case with dummy
models. Injury in the sense of biological damage does not exist in todays mechanical
dummies, because dummy devices are designed for multiple re-use without repair. The
danger of injury to humans is deduced indirectly from the instrument responses of the
mechanical dummies (or their models), as obtained during a real world (or simulated)
crash test.
Human bone, soft tissue, flesh and organ injury prediction is the primary goal of
impact biomechanics. If injuries can be predicted directly and reliably, then cars can
be designed safer. In impact biomechanics, two classes of human parts and organs may
be distinguished from a purely structural point of view: first the ones who have an
identifiable structural function, and second the ones who have not. Skeletal bones, for
example, are “structural” elements in the sense that they must carry the body weight and
mass, and their material resembles conventional engineering materials. The brain, on
the other hand, has hardly any structural function, and it resembles a tofu-like material
with a maze of reinforcing small and tiny blood vessels, not unlike a soft “composite”
material. The structural response of skeletal bone can be modeled more readily with
standard engineering procedures than the structural response of the brain, and the injury
to skeletal bone can be inferred easily as fracture from its structural response, while
neuronal brain injury is not easily derived from the structural response of the brain
material.
Injury prediction. Injury of human parts, before any healing takes place, can either
be defined as instantaneously irreversible mechanical damage, for example damaged
articulations, broken bones, aorta rupture or soft tissue and organ laceration, or, as a
reduction of the physiological functioning, for example of the neurological functions of
the brain, sometimes without much visible physical damage.
Bone fracture, on the one hand, is largely characterized by the mechanical levels of
stress, strain and rate of strain in the skeletal bones, as calculated readily from accu-
rate mechanical models in the simulation of an impact event. Long bones (femur, ribs,
humerus, etc.), short bones (calcaneous, wrist bones, etc.) and flat bones (skull, pelvis,
scapula) can often be modeled using standard brittle material models for the harder
cortical bone, and standard collapsible foam material models for the softer, spongy, tra-
becular or cancellous bone. Ligaments and tendons, and sometimes passive muscles,
skin, etc., can be modeled fairly well using standard non-linear rubber-like hyper-visco-
elastic materials.
Internal organs have physiological functions. Their structural attachment inside the
body cavities is given by mutual sliding contact, by in and outgoing vessels, by liga-
ments and by sliding contact with the body cavity walls. Their structural response to
impact is harder to calculate and the calculated mechanical response fields are hard to
correlate with their physiological functioning or injury.
The heart can act like a structural vessel, for example, when it is compressed and
shifted in a chest impact. Gross shifts may cause strain and rupture of the aorta, an event
which can be modeled with advanced solid–fluid interaction simulation techniques. The
mechanical simulation of this process requires a detailed model of the heart, the aorta
walls, and of the way the heart and aorta are anchored inside the chest. The blood should
then be modeled as a fluid medium.
The other internal organs are either solid (liver, spleen, kidneys, etc.), hollow (stom-
ach, intestines, bladder, etc.), or spongeous (lungs). The solid organs respond with their
bulk matter to mechanical aggression in crash events. For example, the liver might be
lacerated by the action of a lap belt. However, the tender liver parenchyma is invested
by tough-walled vessels which render the material heterogeneous and anisotropic. The
hollow organs should be modeled as hollow cavities, with an adequate model of their
contents, which might interact with the organ walls during a mechanical aggression.
For the brain, the mechanical stress and strain fields and their histories, once calcu-
lated, must yet be linked to neurological damage. After impact, the neurons are still
there, but they may have ceased to function properly because they became disconnected
TABLE 1.1
AIS injury scale
AIS code Description General injury Thorax injury

(example)
0 No injury – –
1 Minor Abrasions, sprains, cuts, bruises –
2 Moderate Extended abrasions and bruises; extended Single rib fracture

soft tissue wounds; mild brain concussions
without loss of consciousness
3 Serious (not life Open wounds with injuries of vessels and 2–3 rib fractures
threatening) nerves; skull fractures; brain concussions sternum fracture
with loss of consciousness (5–10 minutes)
4 Severe (life threatening; Severe bleeding; multiple fractures with >4 rib fractures
probability of survival) organ damage; brain concussion with 2–3 rib fracture
neurological signs; amputations with hemo/
pneumothorax
5 Critical (survival is Rupture of organs; severe skull and brain >4 rib fractures
uncertain) trauma; epidural and subdural hematoma; with hemo/
unconsciousness over 24 hours pneumothorax
6 Maximum (treatment not Aorta rupture; collapse of thoracic cage; Aorta laceration
possible; virtually brain stem laceration; annular fracture of
unsurvivable) base of skull; separation of the trunk;
destruction of the skull
at certain strain levels, or because these cells were asphyxiated from the pressure gen-
erated by hematomas, which may prevent proper blood supply to uninjured parts of the
brain. While this may or may not create visible mechanical “material” damage, it will
cause reduction or total loss of the brain functions, hence injury.
The definition of biological and medical injury to the internal organs and its cor-
relation with mechanical output fields as obtained from impact biomechanics models
remains an open field for intensive research.
Injury scales. Criteria for injury potential were proposed by G ADD [1961], G ADD
[1966]. The most often used injury scale for impact accidents is the Abbreviated Injury
Scale (AIS). Table 1.1 contains AIS scores and some associated injuries.
References on injury and trauma. Detailed descriptions and further extensive bibli-
ographies of injury and trauma of the skull and facial bone, the brain, the head, the
cervical spine, the thorax, the abdomen, the thoraco-lumbar spine and pelvis and the
extremities can be found in the book by NAHUM and M ELVIN (eds.) [1993] Accidental
Injury – Biomechanics and Prevention.
In this book first general aspects related to impact biomechanics are discussed in the
following chapters: Chapter 1: The Application of Biomechanics to the Understand-
ing of Injury and Healing (F UNG [1993b]); Chapter 2: Instrumentation in Experimental

Design (H ARDY [1993]); Chapter 3: The Use of Public Crash Data in Biomechani-
cal Research (C OMPTON [1993]); Chapter 4: Anthropometric Test Devices (M ERTZ
[1993]); Chapter 5: Radiologic Analysis of Trauma (PATHRIA and R ESNIK [1993]);
Chapter 6: A Review of Mathematical Occupant Simulation Models (P RASAD and
C HOU [1993]); Chapter 7: Development of Crash Injury Protection in Civil Avia-
tion (C HANDLER [1993]); Chapter 8: Occupant Restraint Systems (E PPINGER [1993]);
Chapter 9: Biomechanics of Bone (G OLDSTEIN , F RANKENBURG and K UHN [1993]);
Chapter 10: Biomechanics of Soft Tissues (H AUT [1993]).
It is recommended to read these chapters for obtaining a good background for
the following chapters, which are devoted to the trauma and injury of the individual
body segments: A LLSOP [1993] Skull and Facial Bone Trauma: Experimental aspects
(Chapter 11); M ELVIN , L IGHTHALL and U ENO [1993] Brain Injury Biomechanics
(Chapter 12); N EWMAN [1993] Biomechanics of Head Trauma: Head Protection (Chap-
ter 13); M C E LHANEY and M YERS [1993] Biomechanical Aspects of Cervical Trauma
(Chapter 14); C AVANAUGH [1993] The Biomechanics of Thoracic Trauma (Chap-
ter 15); ROUHANA [1993] Biomechanics of Abdominal Trauma (Chapter 16); K ING
[1993] Injury to the Thoraco–Lumbar Spine and Pelvis (Chapter 17); L EVINE [1993]
Injury to the Extremities (Chapter 18).
These chapters provide a broad overview and many references on injury and trauma
of the human body parts, and most of the brief discussions of injury and trauma in this
article are based on this book. The book further contains chapters on child passenger
protection (Chapter 19), isolated tissue and cellular biomechanics (Chapter 20) and on
vehicle interactions with pedestrians (Chapter 21).
1.4. Human models
Models of mechanical dummies simulate their metallic, rubber and plastic parts. Hu-
man models simulate the response of bone, flesh, muscles, and hollow and solid organs
humans are made of. While humans cannot replace mechanical dummies in real world
crash tests, numerical models of humans can readily replace numerical models of me-
chanical dummies in virtual crash simulations.
Generic and specific models. Depending on the application, human models can be
conceived either as “generic” or as “specific” models.
“Generic” models describe the geometry and the physical properties of average size
members of the population. They are needed for industrial design, whenever objects
are designed for the “average” human user. The average size of the human body can
be expressed in statistical “percentiles” of a given population, where the “nth height
percentile” means that “n” percent of the population is smaller in height. For example,
40% of a population is smaller than its 40th height percentile specimen, while 60%
is taller. Each average height and weight percentile specimen can still have different
relative size distribution of its body segments, as well as different biomechanical prop-
erties. The variations around an average percentile specimen of a population are called
its “stochastic variants”.
“Specific” models describe the geometry and physical properties of given human
subjects. They are needed, for example, for virtual surgery, where the surgeon wants to
deal with the precise bone or organ of a given patient. In the case of generic models, the
acquisition process of the geometry of the body may take time, whereas, for practical
reasons, the time needed for the establishment of specific models must be short. There-
fore, slow mechanical slicing techniques on cadavers can serve for the data acquisition
of generic models, while fast X-ray and scanning techniques on patients are required to
construct specific models.
Generic human models should be comprehensive in the sense that all body sizes,
genders, ages, races and body morph-types are covered. To achieve this goal, great
amounts of anthropometrical and biomechanical data must be acquired and collected
in databases, including for children. In fact, human computer modeling and simulation
created new demands for data which were not needed or collected before, and novel
physical experiments are required. Concerning model validation, modern practice of
simulation tends to reverse the role of physical experiments, or laboratory tests, which
tend to back up model calibration and validation, rather than to yield primary results,
now obtained by the simulations.
Scaling, morphing, aging. Generic models of any type and size should be made avail-
able in data bases and through mathematical scaling, morphing and aging techniques,
which can generate any given percentile human model and its stochastic variants, with
long or short trunks and extremities, thin or fat, older or younger, male or female. Exper-
imental results of standardized validation test cases, together with simulation accuracy
norms, must also be provided in such data bases, allowing the human modeler to judge
the performance and the quality of the models under controlled conditions.
As almost none of these new requirements are met fully today, there is plenty of
room for human model development work. In this article, only the fiftieth percentile
“average” male human models for passenger car occupant safety analysis and design
are discussed in detail. Female and child models are discussed more briefly, since they
are less advanced and their modeling techniques resemble the average male models.
Omnidirectionality. Unlike the well-known existing families of mechanical passenger

car occupant surrogates (“dummies”), which are widely used in standard real crash tests
by the world’s car manufacturers for distinct frontal, side and rear impacts, human mod-
els should not be specialized to certain types of crash. They should rather be modeled as
“omni-directional” objects, to the image of their real counterparts, i.e., respond equally
well for all conceivable types of crash scenarios, impact directions and locations.
1.5. Biomaterials
Biomaterials are “exotic” as compared to most conventional structural materials. A good

starting point for their analysis is nevertheless the existing library of material models,
offered in modern dynamic structural analysis codes, or crash codes. The theory of the
available standard material descriptions can be found in the handbooks of the commer-
cial crash codes and need not to be discussed here in their mathematical detail. Ongoing
work will adapt and refine these existing models as new knowledge on the mechanical
behavior of biomaterials emerges. A condensation of the abundant literature on bioma-
terials is added below in brief discussions on fundamental works on biomaterials.
Basic literature. The older book by Yamada (YAMADA [1970]) contains global infor-
mation about basic material properties of most biological tissues, such as the average
Young’s modulus and the fracture strength of the tested parts and organs, which permits
to get first rough ideas about the mechanical properties of human tissues and organs.
The editor of Yamada’s book, F.G. Evans, states in his 1969 preface:
“. . . It is a unique book in several respects. First, it contains more data on
strength of more tissues from more individuals of different ages than any
other study of which I am aware. Second, all of the material used in the
study was fresh and unembalmed. Third, the tests were made with standard
testing machines of known accuracy or with machines that, after consul-
tation with the manufacturer, had been specifically modified for testing
biological materials. Fourth, all of the human material was obtained from
one ethnic group. Thus the strength characteristics and other mechanical
properties of organs and tissues from Japanese can be easily compared with
those from other racial groups. Fifth, data were included on the strength
characteristics of organs and tissues from other mammals as well as birds,
reptiles, amphibians, and fish.”
These remarks clearly express not only the durable value of this introductory book, but
contain the fundamental specifications for the structure and contents of a comprehen-
sive data base of biological materials. Among the tests that were carried out in order to
characterize the strengths of the materials were tests in tension, compression, bending,
impact bending, impact snapping, torsion, expansion, bursting, tearing, cleavage, shear-
ing, extraction, occlusion, abrasion, crushing and hardness. The book by Yamada next
contains an impressive array of basic information about the mechanical properties for
humans and animals of the loco-motor organs and tissues (bone, cartilage, ligaments,
muscle and tendons); the circulatory organs and tissues (heart, arteries, veins and red
blood cells); the respiratory and digestive organs and tissues (larynx, trachea, lungs,
teeth, masticatory muscles, esophagus, stomach, small intestine, large intestine, liver
and gall bladder); the uro-genital organs and tissues (kidney, ureter, urinary bladder,
uterus, vagina, amnion membrane and umbilical cord); the nervous system, integu-
ment, sense organs and tissues (nerves, dura mater, skin, panniculus adiposus (fat), hair,
nails, horn sheath, cornea and sclera (eye), auricle and tympanic membrane (ear)). The
mechanical properties of certain organs such as the brain, the tongue, the spleen are
missing, however. Then the book compares the mechanical properties of human organs
and tissues according to their strength and with respect to other materials from industry
and nature. Finally, varations (scatter), age effects and aging rates are discussed.
More recently, V IANO [1986] describes the biological structures, material properties
and failure characteristics of bone, articular cartilage, ligament and tendon. In his article,
the load-deformation of biological tissues is presented with particular reference to the
microstructure of the material.
“Although many of the tissues have been characterized as linear, elastic and
isotropic materials, they actually have a more complicated response to load,
which includes stiffening with increasing strain, inelastic yield and strain
rate sensitivity. Failure of compact and cancellous bone depends on the rate,
type and direction of the loading. Soft biological tissues are visco-elastic
and exhibit a higher load tolerance with an increasing rate of loading.”
Viano’s paper includes a discussion on the basic principles of biomechanics and em-
phasizes material properties and failure characteristics of biological tissues subjected
to impact loading. The author presents on more than 30 pages what should be known
from an engineering point of view about biological tissues. He discusses what types
of fibers (collagen), bulky tissue with visco-elastic properties, some of which can con-
solidate (hyaline cartilage), and crystals in bone tissue (calcium), are responsible for
the cohesion of the skeleton (ligaments), the attachment of the muscles to the skeleton
(tendons), the transmission of compression forces across the articular surfaces (artic-
ular cartilage) and for maintaining the overall shape of the skeleton (bones). For each
discussed material the paper describes its biological microstructure and composition,
it discusses laboratory setups for material testing, it gives typical stress-strain samples
and it outlines possible mathematical models to describe the measured properties up to
rupture and fracture.
The textbook by F UNG [1993a] describes “The Mechanical Properties of Living Tis-
sues” (book title). The approach to the description of biomaterials chosen by Fung is
the study of the morphology of the organism, the anatomy of the organ, the histology
of the tissue, and the determination of the mechanical properties of the materials or
tissues in the form of their constitutive equations. The book further deals with setting
up the governing differential or integral equations of biomechanical processes, their
boundary conditions, their calibration, solution and validation on experiments and pre-
dicted results. The constitutive behavior of biomaterials is identified and their equations
are defined for the flow properties of blood, blood cells and their interaction with ves-
sel walls, for bio-visco-elastic fluids, for bio-visco-elastic solids, for blood vessels, for
skeletal muscle (with a description of Hill’s active and passive muscle model), for the
heart muscle, for smooth muscles and for bone, cartilage, tendons and ligaments, includ-
ing the mechanical aspects of the remodeling or growth of certain tissues. The detailed
derivation and the mathematical description of the constitutive equations of living tis-
sues is the most distinguishing feature of Fung’s textbook. For each treated subject, the
book contains extensive lists of references that may be consulted for further reading.
Further collections of biomaterial properties can be found in more recent references,
such as the handbook of biomaterial properties by B LACK and H ASTINGS (eds.) [1998],
which describes in its Part I, from the view point of surgical implants, the properties of
cortical bone, cancellous bone, dentin and enamel, cartilage, fibro-cartilage, ligaments,
tendons and fascia, skin and muscle, brain tissues, arteries, veins and lymphatic vessels,
the intra-ocular lens, blood and related fluids and the vitreous humour. (Part II deals
with the properties of surgical implant materials and Part III with the biocompatibil-
ity of such materials, not relevant in impact biomechanics.) The cortical bone material
is described in its composition (organic, mineral), in its physical properties (density,
electromechanical, other) and in its mechanical properties (dry, wet, scatter within the
skeleton, stiffness, strength, strain rate and visco-elastic effects). At the end of each
chapter additional readings and many references are indicated.
Papers that give detailed stress-strain behavior of biomaterials, including rate ef-
fects, are still scarce. Differences between dead and life tissue behavior are seldom
described, and data are often inaccessible. The natural scatter between tissues from dif-
ferent individuals is sometimes discussed. The numerical analyst is still constrained to
use approximate, incoherent or incomplete data. Many efforts are now undertaken to al-
leviate this lack of data, a need that was generated only recently by the desire to simulate
the biomechanical response the human body using modern computer simulation tools.
Simplest descriptions for biomaterials. Fig. 1.1 shows a selection of some typical bio-
material response curves, as extracted from YANG [1998] (a report of the HUMOS-1
Project, funded by the European Commission under the Industrial and Materials Tech-
nologies program (Brite–EuRam III)). The well-known basic elastic, visco-elastic, and
elasto-plastic material laws that exist in most dynamic codes and have been applied for
biomaterial description. The elastic laws can be linear or nonlinear elastic, isotropic,
orthotropic or hyper-elastic. The linear elastic materials are characterized by the elas-
tic moduli, Poisson’s ratios, the shear moduli and the mass density. The hyper-elastic
materials are characterized by their respective strain energy functions (Mooney, Hart-
Smith, etc.). The elastic-plastic material laws are typically defined with the additional
hardening modulus, the yield strength, the ultimate strength and strain at failure. The
visco-elastic materials need additional coefficients describing the damping, creep and
relaxation behavior. The material laws provided with commercial codes are often suffi-
cient to describe hard tissues, such as long bones.
The application to soft tissues is less evident and more research and tests are needed
to characterize these materials. In particular, the difference between life and dead tissue
behavior is more pronounced in soft than in hard tissues.
In many cases so-called “curve description options” for the standard material laws, as
available in the commercial dynamic codes, can be used in order to encode the results
directly as obtained from biomaterial tests. These options provide a maximum freedom
for the analyst, beyond the usual mathematical descriptions of the materials.
Appendix B gives a summary on the mechanical properties of biomaterials as ex-
tracted mainly from YANG [1998].
Bone materials. Fig. 1.2 exemplifies the most frequently studied bone material. In-
set (a) shows a cross section through the femur head, with the cortical outer shell of
compact bone and the trabecular inner fill of spongeous bone clearly visible. Inset (b)
shows the same basic structure in a cross section through the skull bone. Insets (c) and
(d) (after R IETBERGEN [1996], R IETBERGEN , M ÜLLER , U LRICH , RÜEGSEGGER and
H UISKES [1998] and U LRICH [1998]) show so-called “voxel models” of the bony struc-
ture, where the trabecular structure of the bone is modelled directly in the optical voxel
resolution of micro-scans of the bone. Inset (c), for example, uses several million sim-
plified voxel finite elements to trace (red colour) the linear elastic force path through
the trabeculae from an axial force loading. Inset (d) shows two different voxel densities,
and inset (e) demonstrates that there is practically no visible difference between a real
F IG . 1.1. Some typical biomaterial responses as a function of material type, inter-individual scatter, age, de-
formation rate (compiled by YANG [1998]). (Inset (a): Reproduced by permission of Elsevier Health Sciences
Rights; Insets (b) and (f): Reproduced by permission of The Journal of Bone and Joint Surgery, Inc.; Insets
(c) and (d): Reproduced by permission of Chalmers University of Technology; Inset (e): Reproduced by per-
mission of The Stapp Association; Insets (g) and (h): Reproduced by permission of Lippincott, Williams and
Wilkins.)
F IG . 1.2. Bone material structure and modeling (after R IETBERGEN , M ÜLLER , U LRICH , RÜEGSEGGER
and H UISKES [1998], U LRICH [1998]). (Insets (a)–(c), (e): Reproduced by permission of the Journal of
Biomechanics; Inset (d): Courtesy Dr. Ulrich of ETH Zurich.)
X-ray picture of the bone and the simulated X-ray picture made from a voxel model of
the bone.
This figure stands for the basic need for research concerning the modeling of bio-
materials. The direct modeling of the fine structure of the bone material, used in the
shown example, is certainly the best possible approach to model bone, since it disposes
of uncertain macroscopic averaging processes, but remains in the realm of research
and development. This approach may become common practice, once compute power
will have increased to the required practical levels. Today, bones must be modeled
with “macro” elements, the type of which must be chosen according to the type of
bone (e.g., shells for cortical and solids for trabecular bone), and the material densi-
ties of which must be evaluated approximately from the average local density of the
bone.
Similar remarks can be made for other types of biomaterials, which must be investi-
gated indirectly using “smeared” properties and macro modeling techniques. The brain
material, for example, is modeled as solids with homogenized gray or white matter,
without taking into account the system of very fine blood vessels that it contains. If this
system could be modeled in detail, injury to the vascular system of the brain could be ac-
cessed directly, see first attempts made in Fig. 3.3(e)–(h). Today brain injury is accessed
indirectly by correlating mechanical field variables to injury through calibration.
References on biomaterial tests, laws, models and simulation. The literature on bio-
materials and related subjects is relatively abundant. Appendix B contains references
on the biomaterials While many of the indicated references deal with the experimental
evaluation of biomaterial properties, others deal with the aspects of their modeling, the
use of these materials in biomechanical models and the characterization of trauma and
injury.
1.6. Human model validation

Segment and whole body validation. Provided human models can be built and the
biomaterials can be calibrated, one of the greatest challenges remains their proper val-
idation. A considerable number of tests on cadaver body segments, whole cadaver
bodies and life volunteers were performed in the past, e.g., as listed in the report of
the HUMOS-1 project: “Validation Data Base”, ROBIN [1999]. As discussed in a later
section, the European HUMOS-1 project (Human Models for Safety) produced a first
near 50th percentile male European human model in a project funded by the European
Commission (HUMOS-1, 1999–2001; HUMOS-2 is under way). The tests listed in this
reference comprise the following topics, Fig. 1.3:
Head/neck complex: Frontal tests at 15 g, inset (a); lateral tests at 7 g, inset (b);
oblique tests at 10 g, after E WING , T HOMAS , L USTICK , M UZZY III, W ILLEMS and
M AJEWSKI [1976] (not shown).
Thorax frontal impact: frontal impactor tests by INRETS at low and high velocity (not
shown); frontal impactor tests by K ROELL , S CHNEIDER and NAHUM [1971], K ROELL ,
S CHNEIDER and NAHUM [1974] at 4.9 m/s, at 6.7 m/s and at 9.9 m/s, inset (c); at
7.0 m/s with seat back (not shown); frontal thorax impactor tests by S TALNAKER ,
M C E LHANEY, ROBERTS and T ROLLOPE [1973] (not shown).
Thorax belt compression tests: by C ESARI and B OUQUET [1990], C ESARI and B OU -
QUET [1994], with a 22.4 kg mass at 2.9 m/s and 7.8 m/s impact velocities, with a
76.1 kg mass at 2.9 m/s impact velocity, inset (d).
Thorax lateral impact: lateral impactor tests by INRETS on the thorax at 3.3 m/s, at
5.9 m/s, inset (e).
Thorax oblique impact: oblique thorax impactor tests by V IANO [1989] at 4.42 m/s,
at 6.52 m/s, at 9.32 m/s, inset (f).
Abdomen impact tests: frontal impactor tests on the abdomen by C AVANAUGH ,
N YQUIST, G OLDBERG and K ING [1986], inset (g); oblique impactor tests by Ca-
vanaugh at 31.4 kg and 6.9 m/s (not shown); oblique impactor tests by Viano on the
abdomen at 4.8 m/s, at 6.8 m/s, at 9.4 m/s (not shown).
Pelvis impact tests: lateral tests on the pelvis by INRETS at 3.35 m/s, at 6.6 m/s,
inset (h); lateral impact tests by Viano at 5.2 m/s, at 9.8 m/s (not shown) (B OUQUET,
R AMET, B ERMOND and C ESARI [1994]).
It is clear that many more tests must be done in order to capture the biomechanical
characteristics of humans, not only of the “average” subject (50th percentile male), but
F IG . 1.3. Some typical test results for the validation of human models (compiled by ROBIN [1999]). (Insets
(a)–(d), (f) and (g): Reproduced by permission of The Stapp Association; Inset (e): Reproduced by permission
of INRETS; Inset (h): Material in the public domain by U.S. Department of Transportation.)
also of the inter-individual dispersions that distinguish humans. In order to arrive at

“average” responses, and at their likely dispersions, multiple tests should be performed
and the results collected in “corridors”, which represent best the variable “response”
of a given class of human individuals, such as, for example, the 50th percentile male.
Some pertinent test results are summarized in Fig. 1.3.
In particular, whole body cadaver sled tests permit to assess the overall response of
humans in car accident scenarios. Due to the fact that human models can now be built,
calibrated, validated and used in crash simulation or in virtual crash testing, there is a
pressing need for reliable data. The necessary tests on human volunteers and cadavers
are subject to severe ethical control, which considerably restricts the frequency and
number by which such tests can be performed.
2. Overview on mechanical dummies and models
This section may be skipped by readers not interested in mechanical dummies. The
material is provided for to give an overview on the mechanical occupant surrogates or
legal crash “dummies” as presently used by the auto industry for certification of new
car models (K ISIELEWICZ and A NDOH [1994]).
Mechanical dummies (occupant surrogates) and their numerical models are used
heavily in crash tests and numerical simulations for safe car design. Due to the large
number of car crashes each year, crash tests are administered by the National Highway
Traffic Safety Administration (NHTSA), an agency within the United States Depart-
ment of Transportation (DoT). About 35 new model cars have been tested every year
since 1979 under the New Car Assessment Program (NCAP). The tests are to see how
well different vehicles protect front-seat passengers in a car-to-car head-on collision at
equal speeds. The head-on collision is used instead of a rear or a side collision because
this is the collision that causes the most deaths and injuries.
The US federal law requires all cars to pass a 30 mph frontal rigid barrier test, so
NCAP crash tests on fixed barriers (rigid walls) are performed at 35 mph (56.3 km/h),
which corresponds to an impact of two identical cars colliding head-on at a relative ve-
locity of 70 mph (112.6 km/h). These tests show the difference in protection in different
car models. The results of the crashes are given on a one-to five star rating, with five
being the highest level of protection.
These crash tests are all administered with dummies, the dummies are always wearing
seat belts because they are standard equipment on cars today, air bags are used whenever
they are available, and test results are only useful in comparing cars of similar weight
(within 500 pounds (227 kg) of each other). Dummies heads and knees are painted
before a test to see where these areas of the body make contact with the car.
The Hybrid III dummy family is used today in frontal impact tests. For side impact
tests, the US DoT SID and the EuroSID special side impact dummies are used. Some
of these mechanical occupant surrogates and typical protective measures are shown in
Fig. 2.1.
The Hybrid III dummy family. The 50th percentile Hybrid III represents a man of
average size. The European standard 50th percentile man is assumed 1.75 meters tall
and having a body weight of 75.5 kilograms. Different “standard” sizes may exist in
F IG . 2.1. Front impact dummies and passive safety systems. (Inset (a): Reproduced by permission of First
Technology Safety Systems, Inc.; Inset (b) lower right-hand side: Courtesy Autoliv, BMW (F OGRASCHER
[1998]); Inset (b) lower left-hand side: Courtesy AUDI.)
TABLE 2.1
Average body heights and weights (subject to variations)
Percentile age Height Weight

[cm] [foot in ] [kg] [lbs]
50th Hybrid III dummy 178 ∼ 5 10 77.11 170

50th European adult male 175 ∼ 5 9 75.5 166.45
5th adult female 152 ∼ 5 49.89 110
95th adult male 188 ∼ 6 2 101.15 223

6 year old child 113 ∼ 3 8 12 21.32 47
3 year old child 99
∼3 3 14.97 33
different countries. Born in the USA in the labs of General Motors, the 50th percentile
Hybrid III is the standard dummy used in frontal crash tests all over the world. It is
called a hybrid, because it was created by combining parts of two different types of
dummies. Beside the 50th percentile male, there are the 5th percentile female, the 95th
percentile male and the 6 year-old and 3 year-old child dummies, Table 2.1.
Hybrid III dummy models. Fig. 2.2 shows numerical models of members of the Hy-
brid III dummy family (after FTSS/ESI Software). The models shown under Fig. 2.2(a)
have 25 878 (50th percentile), 24 316 (5th percentile), 27 872 (95th percentile), 34 535
(6 year old child) and 13 345 (3 year old child) deformable finite elements, respectively.
In (b) a Hybrid III dummy model is shown in a driver position. The 50th percentile male
Hybrid III model is shown in Fig. 2.2(c) through (e). Insets (c) and (d) are finite ele-
ment models, while inset (e) is a section through the simpler multi-body version. Most
dummy models are made either as simpler multi-body models, or as more detailed finite
element models.
F IG . 2.2. Numerical models of the Hybrid III dummy family (FTSS/ESI Software). (Reproduced by permis-
sion of First Technology Safety Systems, Inc.)
Multi-body modelling techniques comprise linked rigid body tree structures, which
contain relatively few deformable parts and which are linked together at the intersec-
tions of their anatomical segments. They execute faster but cannot yield detailed injury
data. Finite element models are made of the usual standard library of finite elements of
the used crash codes (solids, shells, membranes, beams, bars, springs, etc). They take
more central processor unit (CPU) computer time, but can yield response data, which
are more readily linked to human injuries. Typical solver codes used to analyse car crash
scenarios execute the explicit time integration scheme for the set of non-linear equations
of motion in the nodal degrees of freedom.
Standard injury criteria. The standard way of assessing injuries of vehicle occupants
are heuristic injury coefficients that are calculated from injury criteria defined from the
instrumented front or side impact dummy responses in crash tests.
Head injuries of occupants are assessed from the “Head Injury Coefficient” (HIC),
t2 2.5
1
HIC = max (t2 − t1 ) a(t) dt ,
t1 <t2 t2 − t1 t1
where t = t2 t t1 is a normed time window (e.g., 22.5 ms) that is shifted along
the Hybrid III head acceleration magnitude time history, a(t), as recorded by the head
accelerometers, to find the maximal value of the HIC coefficient over the duration of the
crash event. If the calculated HIC-value is below critical (e.g., 1000), then it is assumed
that no serious injury (skull fracture; neuro-vascular damage) occurs.
Neck injuries can be assessed (among other criteria) from the Nij neck injury crite-
rion,
Nij = FZ /FZ,crit + MY /MY,crit ,
where FZ is the recorded neck axial force (tension/compression), MY is the recorded
sagittal neck bending moment (flexion/extension) and subscripts “crit” indicate the re-
spective critical values, set such that Nij = 1.0 corresponds to a 30% probability of
injury.
Similar criteria exist for the thorax (chest acceleration, chest compression, viscous
criterion, side impact dummy rib deflection, thoracic trauma index), for the abdomen
(abdominal peak force, pelvis acceleration, pubic symphysis peak force) and for the
lower extremity (femur load, tibia index).
SID, EuroSID, BioSID, SID II(s). Hybrid III dummies are designed to be used in
frontal crash tests. For tests representing crashes in which a vehicle is struck on the
side, a number of dedicated side-impact dummies have been created to measure injury
risk to the ribs, spine, and internal organs, such as the liver and spleen, Fig. 2.3.
US DoT SID was the first side-impact dummy. It was developed in the late 1970s
by the US National Highway Traffic Safety Administration (NHTSA) of the US De-
partment of Transportation (DoT) and is used in US government-required side-impact
testing of new cars.
EuroSID was developed by the European Experimental Vehicles Committee (EEVC)
and is used to assess compliance with the European side-impact requirements.
BioSID is based on a General Motors design. It is more advanced than SID and
EuroSID, but it is not specified as a test dummy to be used in legal tests.
SID, EuroSID, and BioSID are designed to represent 50th percentile or average-size
men 5 feet 10 inches (1.78 m) tall and 170 pounds (77.11 kg).
F IG . 2.3. Side impact dummies (hardware). (a) US DoT SID; (b) EuroSID; (c) BioSID; (d) SID II(s) 5th
percentile female. (Reproduced by permission of First Technology Safety Systems, Inc.)
F IG . 2.4. US DoT SID and EuroSID side impact dummy models (FAT/ESI Software).
SID II(s) represents a 5th percentile small female who is 5 feet (1.52 m) tall and
weighs 110 pounds (49.89 kg). SID II(s) was created by a research partnership of US
automakers. It is the first in a family of technologically advanced side-impact dummies.
SID measures the acceleration of the spine and ribs. Acceleration is the rate of veloc-
ity change, and measuring it indicates the forces inflicted on the body during the crash.
EuroSID, BioSID, and SID II(s) measure acceleration plus compression of the rib cage.
Compression refers to the extent body regions are squeezed during the impact and is
used as an indicator of injury to internal organs.
Side impact dummy models. In Fig. 2.4 finite element models of the US DoT SID
and of the EuroSID dummy are shown. These models were elaborated on the basis of
material, component and whole body tests, performed by the German car manufacturer
consortium FAT. The models have about 38 000 deformable finite elements and a time
step of 1.5 microseconds in explicit solver codes.
BioRID. A rear-impact dummy has been developed to measure the risk of minor neck
injuries, sometimes called whiplash, in low-speed rear-end crashes, which is a big prob-
lem worldwide.
BioRID was developed in the late 1990s by a consortium of Chalmers University of
Technology in Sweden, restraint manufacturer Autoliv, and automakers Saab and Volvo.
It is designed to represent a 50th percentile or average-size man, 5 feet 10 inches tall and
170 pounds in weight, Fig. 2.5, DAVIDSSON , F LOGARD , L ÖVSUND and S VENSSON
[1999].
BioRID has been designed especially to study the relative motion of the head and
torso. For tests representing crashes in which a vehicle is struck in the rear, BioRID
can help researchers learn more about how seatbacks, head restraints, and other vehicle
characteristics influence the likelihood of whiplash injury.
Unlike Hybrid III dummies, the BioRID spine is composed of 24 vertebra-like seg-
ments, so that in a rear-end crash BioRID interacts with vehicle seats and head restraints
in a more humanlike way than the Hybrid III. The BioRID segmented neck can take on
the same shapes observed in human necks during rear-end collisions, an important char-
acteristic for measuring some risk factors associated with whiplash injury. Comparative
cross sections through the human spine and BioRID are shown in Fig. 2.5.
F IG . 2.5. BioRID Rear Impact Dummy and comparative human section. (Reproduced by permission of
The Stapp Association.)
F IG . 2.6. Child Restraint Air Bag Interaction dummy (CRABI). (Reproduced by permission of First Tech-
nology Safety Systems, Inc.)
CRABI. The Child Restraint Air Bag Interaction dummy was developed by First Tech-
nology Safety Systems (FTSS) to represent children, Fig. 2.6. It is used to evaluate child
restraint systems, including airbags. There are three sizes: 18 month-old, 12 month-old,
and 6 month-old. These dummies have sensors in the head, neck, chest, back, and pelvis
that measure forces and accelerations.
THOR. This advanced 50th percentile male dummy is being developed in the United
States by NHTSA for use in frontal crash tests, Fig. 2.7 (http://www-nrd.nhtsa.dot.gov/
departments/nrd-51/THORAdv/THORAdv.htm). THOR has more human-like features
than Hybrid III, including a spine and pelvis that allow the dummy to assume various
seating positions, such as slouching, for example, or sitting upright. THOR also has
sensors in its face that measure forces so that the risk of facial injury can be assessed,
which is not possible with current dummies. In fact, all of THORs standard sensors will
provide more injury measurements than those available on Hybrid III.
F IG . 2.7. Advanced THOR dummy. (Reproduced by permission of NHTSA, U.S. Government.)
3. Overview of existing human models for occupant safety
The following paragraphs briefly outline the present state of the art in human biome-
chanical modeling for occupant safety. Most of the information has been drawn from
the indicated INTERNET web sites. The interested reader is invited to consult these
sites, and more, to get up-to-date information of this rapidly expanding field. The se-
lected examples demonstrate the extensive level of development of human models for
impact biomechanics by research institutions and by private industry. The fact that car
manufacturers invest actively in human models reflects the need for the use of human
models in safe car design. At the same time a unification of these efforts is needed, as
reflected, for example, by the pre-competitive European joint development project HU-
MOS, in which five car manufacturers, several equipment suppliers, several research
institutes and three engineering software houses are active partners. While the official
approval by the respective National Road Safety Administration authorities for the use
of human models for safe car design in “legal virtual crash tests” is pending, new car
models are being certified using mechanical dummies in real physical legal crash tests.
An interesting concept on the way towards legal virtual crash tests with human models
has been introduced recently by the National Highway Traffic Safety Administration
(NHTSA), US Department of Transportation (DoT), which can be considered an en-
couraging step towards the increased use of human models in safe car design. This new
concept is discussed first.
3.1. SIMon (Simulated Injury Monitor)
NHTSA experts have developed human models (http://www.nhtsa.dot.gov), which di-

rectly simulate bodily injury, unlike dummies or models of dummies, which access
injury through equivalent measures. The developed models are meant to help new, ad-
vanced, mechanical dummy design (e.g., THOR) on the one hand, and, on the other
hand, they serve as vehicles in the numerical interpretation of the enriched output data,
harvested from the new generation dummies in crash tests.
SIMon-Head. The recently released first SIMon-Head model is discussed below and
it provides the first step towards a new standardized analytical occupant safety analysis
methodology, called SIMon (Simulated Injury Monitor). The objective of this particular
research is to evaluate injury to the soft tissue of the human brain using finite element
models of the brain together with dynamic load data from mechanical dummies, har-
vested in actual crash tests. The so far released SIMon-Head model consists in a CD
Rom with an NT software package that can accept the measured output of nascent new
generation mechanical dummies (e.g., THOR). The new dummy head is equipped with
nine instead of three accelerometers, which permits to record translational, as well as ro-
tational accelerations of the head. The SIMon-Head package uses these comprehensive
acceleration time histories as an input to a built-in calibrated finite element model of
the head and brain, to which it applies the recorded accelerations by running an explicit
solver code, itself locked into the SIMon-Head package.
The package then analyzes the built-in head/brain model output data and it generates
three new brain injury coefficients, namely the “Cumulated Strain Damage Measure
(CSDM)”, the “Dilatation Damage Measure (DDM)” and the “Relative Motion Damage
Measure (RMDM)”. The CSDM tells what cumulative volume fraction of the brain
matter experienced at some time principal strains larger than a fixed threshold value
(15%), known to cause Diffuse Axonal Injury (DAI). The DDM tells what instantaneous
volume fraction of the brain matter experienced negative dynamic pressures that can
cause vaporization of the cerebral fluids, and contusion. The RMDM tells the percentage
of the bridging veins that have stretched beyond a limit curve in a calibrated strain vs.
strain-rate diagram, each possibly causing Acute Subdural Hematoma (ASDH) through
rupture. The bridging veins connect the soft brain tissues to the skull and may rupture
through excessive shearing motions of the brain with respect to the skull.
This new concept will raise the level of precision for injury prediction by directly
addressing different types of local injuries, rather than by comparing abstract global
coefficients, such as the well-known Head Injury Coefficient (HIC), with calibrated
threshold values (cf. NHTSA Federal Motor Vehicle Safety Standard (FMVSS)). The
HIC is an integral of translational dummy head accelerations over a moving fixed time
window, which tells the maximum acceleration the head experienced over the fixed time
interval during the crash event. This measure can be linked to real brain injuries only
in a purely heuristic fashion. The SIMon concept will be extended in the near future to
further body segments (neck, thorax, femur, etc.), and it will undoubtedly accelerate the
widespread use of human models in safe car design.
Fig. 3.1(a)–(h) show a typical (side) impact experiment in a crash laboratory (a), a
model of the advanced (frontal) crash test dummy THOR (b), as well as several views
of the SIMon Head model, labeled: a mid-coronal section view (c), a coronal-sagittal
view, highlighting the boundary between the falx cerebri and the skull (d), a 3D view of
the opened SIMon skull model (e), a 3D top view of the brain (f), a top view indicating
the location of the parasagittal bridging veins (g), and a 3D view of a brain model (h).
SIMon is designed to provide head (Fig. 3.1), and later neck and thorax (Fig. 3.2),
and lower extremity models of occupants, including women and children. These models
of human parts can be driven by instrumentation data from advanced crash dummies.
F IG . 3.1. SIMon Head and Brain Models (http://www.nhtsa.dot.gov). (Reproduced by permission of

NHTSA, U.S. Government.)
F IG . 3.2. NHTSA Neck and Chest Models (http://www.nhtsa.dot.gov). (Reproduced by permission of

NHTSA, U.S. Government.)
References for SIMon-Head. The following references may be consulted on the sub-
ject of the SIMon-Head model: A L -B SHARAT, H ARDY, YANG , K HALIL , TASHMAN
and K ING [1999] on brain/skull relative motions; BANDAK and E PPINGER [1994],
BANDAK , TANNOUS , Z HANG , T ORIDIS and E PPINGER [1996], BANDAK , TANNOUS ,
Z HANG , D I M ASI , M ASIELLO and E PPINGER [2001] on brain FE models and SIMon
Head; FALLENSTEIN , H ULCE and M ELVIN [1970] on dynamic mechanical prop-
erties of human brain tissue; G ENNARELLI and T HIBAULT [1982], G ENNARELLI ,
T HIBAULT, T OMEI , W ISER , G RAHAM and A DAMS [1987] on biomechanics of acute
subdural hematoma (1982a) and on directional dependence of axonal brain injury
(1987); L OWENHIELM [1974] on dynamic properties of bridging veins; M ARGULIES
and T HIBAULT [1992] on diffuse axonal injury tolerance criteria; M EANY, S MITH ,
ROSS and G ENNARELLI [1993] on diffuse axonal threshold injury animal tests;
N USHOLTZ , W ILEY and G LASCOE [1995] on cavitation effects in head impact model;
O MMAYA and H IRSCH [1971] on cerebral concussion tolerance in primates; O NO ,
K IKUCHI , NAKAMURA , KOBAYASHI and NAKAMURA [1980] on head injury toler-
ance for sagittal impact from tests.
3.2. Wayne State University human models
For over sixty years, the Wayne State University (WSU) Bioengineering Center has pi-
oneered biomechanics research and issued injury tolerance thresholds. Over the past
years, the center engaged in a continued activity of the development of models of the
human body and its parts, in particular the WSU brain and head model, as shown be-
low. These models are among the most advanced, and their validation is substantiated
by experiments performed at the center itself. The WSU human models have served
many workers and institutions as a basis for their own development and research (Ford,
General Motors, Nissan, Toyota, ESI, Mecalog, etc.).
Fig. 3.3 gives an overview on the WSU human models. The reported information
can be found on their web site http://ttb.eng.wayne.edu, as well as in publications by
Z HANG , YANG , DWARAMPUDI , O MORI , L I , C HANG , H ARDY, K HALIL and K ING
[2001] and H ARDY, F OSTER , M ASON , YANG , K ING and TASHMAN [2001] for the
WSU head injury model, Fig. 3.3(a)–(d); in Z HANG , BAE , H ARDY, M ONSON , M AN -
LEY, G OLDSMITH , YANG and K ING [2002] for the WSU vascular brain model,
Fig. 3.3(e)–(h); in YANG , Z HU , L UAN , Z HAO and B EGEMAN [1998] for the WSU
neck model, Fig. 3.3(i), (j); after S HAH , YANG , H ARDY, WANG and K ING [2001] for
the WSU chest model, Fig. 3.3(k)–(m); in L EE and YANG [2001] for the WSU abdomen
model, Fig. 3.3(n)–(q).
3.3. THUMS (Total Human Model for Safety)
The Total Human Model for Safety (THUMS) was recently assembled and tested
by Toyota Research Company, see Fig. 3.4(a)–(f), cf. F URUKAWA , F URUSU and
M IKI [2002], I WAMOTO , K ISANUKI , WATANABE , F URUSU , M IKI and H ASEGAWA
[2002], K IMPARA , I WAMOTO and M IKI [2002], M AENO and H ASEGAWA [2001],
NAGASAKA , I WAMOTO , M IZUNO , M IKI and H ASEGAWA [2002], O SHITA , O MORI ,
F IG . 3.3. The WSU human models. (Insets (a)–(h), (j) and (n)–(q): Reproduced by permission of The Stapp
Association; Insets (i) and (k)–(m): Reproduced by permission of King H. Yang, Wayne State University Bio-
engineering Center; Insets (a)–(d): WSUHIM Head Injury Model (Z HANG , YANG , D WARAMPUDI , O MORI ,
L I , C HANG , H ARDY, K HALIL and K ING [2001]) and response under occipital impact test C755-T2 (H ARDY,
F OSTER , M ASON , YANG , K ING and TASHMAN [2001]); Insets (e)–(h): WSU 2D Vasculated Brain Injury
Model (Z HANG , BAE , H ARDY, M ONSON , M ANLEY, G OLDSMITH , YANG and K ING [2002]) and response
under impact test nb.37 (N AHUM , S MITH and WARD [1977]); Insets (i) and (j): WSU Neck Model structure
and response under whiplash conditions (YANG , Z HU , L UAN , Z HAO and B EGEMAN [1998]); Insets (k) to
(q): WSU chest model (k), (l), (m) (S HAH , YANG , H ARDY, WANG and K ING [2001]), abdomen model (n),
(o) (L EE and YANG [2001]) and response (p), (q) under lateral pendulum impact (V IANO [1989]).)
F IG . 3.3. (Continued.)
NAKAHIRA and M IKI [2002] and WATANABE , I SHIHARA , F URUSU , K ATO and M IKI
[2001] and their web site http://www.tytlabs.co.jp/eindex.html. This considerable effort
reflects the urgent need for car industry using human models for safe car design. A first
50th percentile human male model was completed in 2000, based on their own devel-
opment and on models from Wayne State University. This model is relatively detailed,
since it comprises more than 80 000 elements, which is about three times the density
of the HUMOS-1 model, as discussed in a separate section of this article. Since con-
ventional mechanical crash dummy models often have the same level of refinement, the
model is suitable for running crash simulations.
THUMS is a family of models, which comprises the AM50 50th percentile male,
Fig. 3.4(a), the AF05 5th percentile female and the 6 year old child, (b), and a pedestrian
model, (c). The internal organs of the AM50 model are shown in inset (d). Insets (e) and
(f) show the deformed shapes of the AM50 and the AF05 models, respectively, under
the Kroell chest pendulum impact tests, with a pendulum mass of 23.4 kg, a diameter
of 150 mm and an impact velocity of 7.29 m/s. The simulation results were compared
with the Kroell tests (K ROELL [1971] and K ROELL , S CHNEIDER and NAHUM [1971]).
The pedestrian simulations with 40 km/h bending and shearing tests, inset (c), were
compared to tests by K AJZER , S CHROEDER , I SHIKAWA , M ATSUI and B OSCH [1997].
3.4. LAB human model
The LAB (Laboratoire d’Accidentologie et de Biomécanique of PSA Peugeot, Citroën,

RENAULT), in collaboration with CEESAR, ENSAM and INRETS, have developed
a complete 50th percentile male human finite element model with 10 000 elements,
Fig. 3.5(a)–(i).
The material properties were taken from the literature and a large data base of 30
test configurations and 120 test corridors was compiled and used to validate the model.
Comparative studies were performed concerning the differences in the behavior of hu-
man models as compared to dummy models in frontal and lateral impact conditions.
Fig. 3.5 gives an overview on this pioneering model and its comparisons with models
of front impact (HYBRID III) and side impact (EuroSID) dummy models.
F IG . 3.4. TOYOTA’s THUMS Total Model for Human Safety family. (Reproduced by permission of Kazuo
Miki, Toyota CRDL.)
F IG . 3.5. The LAB human and dummy models (BAUDRIT, H AMON , S ONG , ROBIN and L E C OZ [1999],
L IZEE , ROBIN , S ONG , B ERTHOLON , L E C OZ , B ESNAULT and L AVASTE [1998]). (Reproduced by permis-
sion of The Stapp Association.)
References for the LAB model. The LAB model has been published in the following
papers: BAUDRIT, H AMON , S ONG , ROBIN and L E C OZ [1999] on comparing dummy
and human models in frontal and lateral impacts; L IZEE , ROBIN , S ONG , B ERTHOLON ,
L E C OZ , B ESNAULT and L AVASTE [1998] on the development of a 3D FE human body
F IG . 3.6. Madymo multi-body and deformable finite element models of the human body. (Reproduced by
permission of TNO Automotive.)
model; W ILLINGER , K ANG and D IAW [1999] on the validation of a 3D FE human

model against experimental impacts.
3.5. MADYMO human models
MADYMO is a TNO Automotive engineering software tool that is used for the design
of occupant safety systems. The following extract on their human models is drawn from
their web site, http://www.madymo.com, Fig. 3.6(a)–(f).
Human body models have been developed for TNOs software program MADYMO,
using a modular approach. Several combinations of detailed multi-body and finite el-
ement (FE) segment models are available. The models have been validated for impact
F IG . 3.7. Radioss human models (http://www.radioss.com). (Insets (a)–(d): Reproduced by permission of

MECALOG Sarl; Insets (e) and (f): Reproduced by permission of The Stapp Association.)
loading. For their multi-body models, a combination of modeling techniques was ap-
plied using rigid bodies for most segments, but describing the thorax as a deformable
structure.
A finite element mesh of the entire human body has been developed by TNO, based
on the data produced in the EU project HUMOS, Fig. 3.6(g). The mesh was based
on a European 50th percentile human in a seated driving position with a detailed 3-D
numerical description of the subject’s geometry. Apart of this full body model, TNO
offers a series of deformable models of various body parts.
3.6. RADIOSS human models
As TNO and ESI Software, Mecalog is a partner in the European HUMOS projects.
They have developed their own encrypted model from the common HUMOS-1 project
data base. Further models of human parts were developed by Mecalog, as shown in
Fig. 3.7(a)–(f). Their head and skull model was developed in collaboration with the Uni-
versity of Strassbourg (W ILLINGER , K ANG and D IAW [1999]). Their lower extremity
model was developed in collaboration with WSU (B EILLAS , L AVASTE , N ICOLOPOU -
LOS , K AYVANTASH , YANG and ROBIN [1999], B EILLAS , B EGEMAN , YANG , K ING ,
A RNOUX , K ANG , K AYVANTASH , B RUNET, C AVALLERO and P RASAD [2001]). More
information can be found on their web site http://www.radioss.com.
C HAPTER II
Human Articulated Multi-Body Models

The numerical models and materials presented in this chapter are based on work carried
out at ESI Software and the University of West Bohemia (Robby family), and at IPS
International and Hong-Ik University (H-Models).
4. Human Articulated Rigid Body (HARB) models
Open tree chain structure. The kinematics of the human body can be described to first
order accuracy by the kinematics of a chain of articulated rigid bodies. Such models are
computationally efficient, but they provide only limited information. Each member of
the human body can be represented as a rigid body, while each skeletal joint is modeled
by a corresponding numerical joint element or using non-linear springs. Modern crash
codes have multi-body linkage algorithms with a comprehensive set of joint models, by
which open-tree linked rigid body structures can be treated effectively. Details about
these modeling techniques can be found in the handbooks of these codes. Human Artic-
ulated Rigid Body (HARB) models have therefore been introduced, which can describe
the overall kinematics of the model under a crash load scenario. Fig. 4.1 shows a fifti-
eth percentile male HARB model, named “Robby”, with its exterior skin, its tree link
structure and its skeletal structure (ROBBY 1 [1997], ROBBY 2 [1998], H YNCIK [1997],
H YNCIK [1999a], H YNCIK [2001a], H YNCIK [2002a], H AUG , B EAUGONIN , T RAME -
CON and H YNCIK [1999], B EAUGONIN , H AUG and H YNCIK [1998]).
Inset (a) of Fig. 4.1 defines the “sagittal”, “coronal” (or “frontal”) and “transverse”
(or “horizontal”) planes, used in anatomy to situate the parts of the body. To situate
body parts relative to the body, directions are convened as follows: “anterior” pointing
towards the front of the body, “posterior” towards the back, “medial” towards the mid-
line, “lateral” away from the midline, “proximal” closer to the trunk and “distal” away
from the trunk.
In inset (b), the upright HYBRID III 50th percentile mechanical dummy model is
superimposed to the human model Inset (c) shows the upright skeleton with joints. In-
sets (d) to (f) show the model placed into the posture of a driver. Inset (g) details the
articulated spine of the model.
Model geometry. The geometry of the shown fiftieth percentile male HARB model is
based on the anatomical data sets available from DIGIMATION (D IGIMATION /V IEW -
POINT C ATALOG [2002]) and the V ISIBLE H UMAN P ROJECT [1994]. The exterior skin
271
272 E. Haug et al. C HAPTER II
F IG . 4.1. Human Articulated Rigid Body (HARB) model: 50 percentile male (ROBBY 1 [1997], ROBBY 2
[1998], H YNCIK [1997], H YNCIK [1999a], H YNCIK [2001a], H YNCIK [2002a], H AUG , B EAUGONIN ,
T RAMECON and H YNCIK [1999], B EAUGONIN , H AUG and H YNCIK [1998]).
S ECTION 4 Human Articulated Multi-Body Models 273
and the skeleton of this particular HARB version are modeled with mostly triangular
and quadrilateral rigid facets, while the anatomical joints are modeled with nonlinear
joint elements that translate the major degrees of freedom and stiffness properties of the
skeletal articulations. The body segments head, neck, thorax, upper arms, lower arms,
hands, pelvis-abdomen, thighs, lower legs and feet represent each a rigid body made of
its skin and skeletal parts (hand and foot articulations beyond the wrist and ankle joints
are usually not modeled for crash simulations).
Mass and inertia properties. Each segment has the center of gravity, mass, principal
inertias and directions of principal inertias of a fiftieth percentile male, as given in the
pivotal UMTRI reports (ROBBINS [1983]), where the anthropometric, the mass and
inertia and the basic joint biomechanical properties of the average fiftieth percentile
American citizen and of the fifth percentile female and the ninety-fifth percentile male
have been collected from an extensive study on human cadavers. The results of this
report have led to the construction of the well known Hybrid III mechanical dummy
family. One family of the discussed human HARB models is named after the reports by
Robbins (Robby, Robina, Bobby).
Contact definition. The modeled body segments can interact with themselves and with
the environmental structures and obstacles via their built in and user defined contact in-
terface definitions. Despite the fact that the external skin is rigid, soft nonlinear contact
penalty spring definitions permit a realistic treatment of the body contacts. Contact al-
gorithms are described in the handbooks of modern solver codes (cf. Appendix A.3).
Joint modeling. The mechanical joint stiffnesses and maximum excursions of their
rotational degrees of freedom are governed by user-specified nonlinear moment-rotation
curves. This modeling technique can also allow for joint failure upon over-extension of
the joints beyond the anatomical motion ranges. The stiffness and resistance properties
are found in the literature (ex: ROBBINS [1983], ROBBINS , S CHNEIDER , S NYDER ,
P FLUG and H AFFNER [1983]) (see also Appendix B). Fig. 4.2 shows typical curves,
based on the Robbins report, for example for the left knee joint.
This model is part of another HARB model, termed H-ARB (Section 9.2). The knee
flexion–extension rotation about the joint r-axis has been modeled approximately by
a moment-rotation curve with negligible stiffness and resistance in the range between
+1.0 and −1.2 radians, and with steep slopes at the ends of the flexion–extension excur-
sion range. The internal–external rotation of the knee joint about the t-axis is modeled
in similar fashion with a reduced angular range. The varus–valgus rotation about the
s-axis is penalized heavily by a stiff slope. The given curves are only approximations
to the real motions of the human knee. In reality, the internal–external rotation range
depends on the flexion–extension angle, there is a certain forward–backward mobility
of the tibia with respect to the femur and the center of rotation of the flexion–extension
rotation is mobile. These additional joint mobilities can be captured with more elaborate
joint models, or when using detailed finite element models, where the articular surfaces
and the stabilizing ligaments and soft tissues are modeled in sufficient detail, which is
beyond the range of simple HARB models.
F IG . 4.2. Typical human joint moment-rotation curves (H-Model).
F IG . 4.3. The attachment of the arm–shoulder complex (ROBBY 2 [1998], H YNCIK [1999a]).
Special attention must be given to the articulations of the arm–shoulder complex.

The only points of the arm–shoulder complex that are fixed with respect to the trunk
are the sterno-clavicular joints, which connect each clavicle to the sternum. The upper
arm is connected to the scapula (shoulder blade) by the gleno-humeral joint, while the
scapula connects to the clavicula by the acromio-clavicular joint. These joints can be
modeled by kinematic joint elements, similar to the one described above. In addition,
each scapula can slide about the outer surface of the rib cage inside pockets, which can
be modeled as a sliding contact interface.
Fig. 4.3 shows the anatomical mobility of the shoulder blade with the extreme up-
down and backward–forward translational motions and inside–outside rotation motions,
inset (a), as well as a sub-system of passive Hill muscle bars that help confine these ex-
treme motions by their passive stretch force reactions, inset (b) (H YNCIK [1999a]). If
these muscles are absent, the motions of the shoulder blade and the clavicula will be
stabilized only by their connecting ligaments, represented in the HARB models by rota-
tional joint elements. A preliminary study by Ludek Hyncik (ROBBY 2 [1998]) showed
that the passive action of the stretched muscles at the ends of the scapular motion ranges
is considerable (more than 30% in average). These muscles therefore help stabilizing the
complex mobility of the arm–shoulder complex.
5. The Hill muscle model
The material described in this section is based on work carried out at ESI Software and
the University of West Bohemia.
Hill-type muscle bars. The mechanical behaviour of the skeletal muscles in the di-
rections of their fibers can be modeled to first order accuracy by Hill-type muscle bars
(H ILL [1970]). Each Hill-type muscle bar element is characterized by the physiologi-
cal cross section area of the muscle, cut perpendicular to the fibers, and by the muscle
fiber stretch and stretch velocity dependent active and passive mechanical properties of
the Hill muscle model, described in Appendix C. The bars cannot, in general, transmit
compressive forces.
These bar type finite elements can be arranged in the directions of the active muscle
fibers. In that approach each skeletal muscle of interest is subdivided into a sufficient
number of segments that can be approximated by bar like elements. For example, each
segment of the biceps muscle of the arm can be approximated for most of its gross ac-
tions by a single bar element, fixed between its anatomical points of origin and insertion
on the skeleton, provided the lever arm topology of the muscle is respected to sufficient
precision.
For interactions with the skeleton and other tissues, and with the environment, further
numerical modeling devices must be introduced, such as contact sliding interfaces and
layers of deformable finite elements. For example, the trapezius muscle, being a flat
surface muscle, can be subdivided into several parallel anatomical segments, each of
which is represented with its tributary anatomical cross section area as one bar element.
In order to facilitate contact of this muscle with the skeleton, each bar element can be
further subdivided along its axis into a number of serial bar elements.
In successive further stages of refinement, skeletal muscles can be modeled as parallel
assemblies of bars, Fig. 5.1(a), (b) (ROBBY-models, H-UE model: cf. Section 9.6); as
bars attached to membrane finite elements which describe the resistance in the direction
perpendicular to the muscle fibers, Fig. 5.1(c) (H-UE model); as heterogeneous two-
dimensional fiber reinforced composite finite shell or membrane finite elements, where
the composite fibers represent the mechanical properties attributed to the active muscle
fibers, and where the composite matrix phase provides the passive in-surface stiffness
and resistance of the muscle; as three-dimensional composite fiber reinforced solid finite
elements, where the composite fiber phase represents the action of the muscle fibers and
where the matrix phase provides the volume bulk response of the body of the muscle. It
can be noted that the discussed muscle models use deformable finite elements, but the
underlying model of the human body can still be a rigid multi-body model, as well as a
deformable finite element model.
F IG . 5.1. Modeling of skeletal muscles with bars and membranes (ROBBY-models, H-UE model).
This underlines the fact that biomechanical models can be held mixed, incorporating
deformable and rigid body models, as required by accuracy and computational effi-
ciency.
The Hill muscle model. The Hill muscle model is one of the simplest phenomenolog-
ical engineering models of the active and passive biomechanical behaviour of skeletal
muscles. Its mathematical description is given in Appendix C.
6. Application of Hill muscle bars
Whiplash simulation. The neck models can be equipped with Hill muscle bars for the
purpose of simulating the effect of the muscle forces on the head/neck displacements
when, for example, a rear car impact victim is bracing the neck muscles in a voluntary
or unvoluntary fashion. Fig. 6.1 shows a head/neck model with a number of Hill muscle
bars added. The corresponding neck “whiplash” event simulation is discussed in the
later section on the H-Neck model (cf. Section 9.4).
Static muscle force distribution. Recently Hill bar models have been used to calculate
the distribution of the skeletal muscle forces of human subjects holding a given set of
static loads at a given fixed posture (H AUG , T RAMECON and A LLAIN [2001]). To this
end, the skeleton of a human model is equipped with Hill muscle bars and the static
loads are applied. Since there are many more muscle segments that can be recruited
to stabilize the kinematic degrees of freedom inherent in the articulated skeleton, the
problem is hyperstatic, i.e., there are more equations than unknowns from which to
calculate the equilibrium muscle forces. In fact there is an infinity of equilibrium force
patterns for which static equilibrium can be achieved.
Optimization of muscle energy. One rational solution to this problem is given by the
assumption that the human body will activate the muscles under a minimum expense
F IG . 6.1. Hill muscle bars added to the H-Neck model for whiplash studies (H-Neck).
of physiological energy. Physiological energy is spent when a muscle must activate to

a constant force level over a given interval of time, even in the absence of any motion
and external mechanical work (“isometric” conditions). This muscle energy is propor-
tional to the product of force and time and the total energy is the sum of all muscle
energies. One solution to the isometric muscle force problem can therefore be found by
minimizing the total active muscle energy. This can be done by solving the associated
optimization problem, where the simplest objective function, f , is given by the square
root of the sum of the squares of the muscle forces, i.e.,
1/2
f= γi (αi − c)2 ,
where the sum ranges over all participating muscle segments, i, activation level c is
the given (average) voluntary level of muscle contraction before the load is applied
(0–100%), αi is the total activation level of the muscle segments that contribute to the
task of carrying the load (0–100%) and γi is a switch that can have value “1” for the
participating groups of muscles and “0”, otherwise. This function can be thought to
express the least possible overall level of muscle activation, or “energy”, to be expended
over a time with constant muscle forces. More complex objective functions have been
proposed in the literature, for example, S EIREG and A RVIKAR [1989].
The constraints for the static optimization process are given by the fact that the ac-
celerations of the links of the kinematic chain, constituted by the involved parts of the
skeleton, must all be equal to zero in a position of static equilibrium. These accelera-
tions can be calculated simply by performing an explicit analysis with a dynamic solver
code, using the relevant muscled skeleton model, under the applied loads. In fact, one
explicit solution time step at time = 0 is enough to determine whether or not the “struc-
ture” is in static equilibrium. At equilibrium, the internal muscle forces must balance
the applied loads, and the accelerations, calculated by the solver at the centers of gravity
of each rigid skeleton link, must be close to or equal to zero.
F IG . 6.2. Hill muscle bars activated at minimal energy under static loads (ROBBY models).
The design parameters of the optimization problem are given by the activation levels,
αi , of the participating muscle segments, which are the result of the optimization.
The bounds on the design parameters are given by 0 < αi < 100%, i.e., the activation
level of a muscle cannot be less than zero and not greater than 100%. The outlined opti-
mization procedure can be solved by standard optimization algorithms and it is applied
to a simple one degree of freedom system.
Example. Fig. 6.2 shows a first example of an arm with fixed shoulder and upper arm,
where a static load is held in place against one elbow rotation degree of freedom. In this
example a minimum of seven muscle segments contributes to the task of supporting the
load, and the only kinematic degree of freedom is the elbow flexion–extension rotation.
Fig. 6.3(a) shows the model of an upper torso with shoulders and arms, equipped
with 132 Hill muscle bars. The model holds a constant static load in the left hand. The
muscle forces are calculated for minimal energy. The colors of the muscle bars range
from blue for zero activation to red for 100% activation.
In Fig. 6.3(b) a symmetric static load is applied to the same model (cubes), which
increases linearly at equal increments. Inset (c) shows the calculated activation levels
of some selected muscle bars at loads ranging from zero to 12 kg in each hand. In
the diagram the agonist muscles (ex. biceps) are seen to quickly raise their levels of
activation, initially in proportion to the load increase, while the less efficient muscles
(ex. latissimus dorsi) initially contribute very little to the task. As soon as the first agonist
muscles become saturated, i.e., reach 100% activation after about 3 kg loads, the less
efficient muscles tend to increase their activation levels faster than linear. After the loads
have reached about 8 kg, the inefficient muscles (ex. latissimus dorsi) must compensate
for the saturation of the agonists and they raise their levels of activation dramatically.
When all muscles reach saturation, the loads can no longer be increased and the limit
load has been reached (2 × 12 kg). Limit loads for different postures are calculated in
the examples shown in Appendix C, paragraph: ‘Applications’, Fig. C.3.
Fig. 6.3(d) and (e) show the model pulling the hand brakes in a passenger car and
holding the steering wheel while driving. In both postures the needed muscle force
F IG . 6.3. HARB muscle forces under different static loads (ROBBY-models).
pattern was calculated for minimal energy. These examples show that articulated rigid
body models can be used together with Hill muscle bars to calculate realistic muscle
force distributions. These forces show the limit loads a subject can hold for a short
time. For long term static muscle loads, such as in driving a car, the forces are less
important, but will lead eventually to muscle fatigue. The levels of fatigue can best be
assessed by knowing the forces that cause the fatigue.
Fatigue. The Hill muscle model can be extended to add muscle fatigue. Under the
isometric conditions, such as in the static scenarios described above, a given muscle
will fatigue when forced to sustain an active muscle force over a prolonged period of
time. If the active force output of a muscle must be constant over time, such as in the
simple examples of holding a given load as shown in Fig. 6.2, then muscle fatigue will
force the muscle to raise its level of activation in order to compensate for the fatigue-
induced drop of force. If the activation level is kept constant, the muscle force will drop
over time. Simple static muscle fatigue models are currently under investigation.
7. Application of HARB models
HARB model families. In Fig. 7.1 models of a fifth percentile female (ROBINA [1998])
and a six year old child (BOBBY – ESI Software) are shown. These models are built
in a fashion which is analogous to the HARB model of the fiftieth percentile male, and
their details are not discussed separately. It must be noted, however, that basic anthro-
pometric data exist mostly for the adult humans, while such data for children are close
to absent. Nevertheless, DIGIMATION (D IGIMATION /V IEWPOINT C ATALOG [2002])
provide the external geometry, the skeleton geometry and the organ geometry of male
and female adults, as well as some muscle surface geometry. The external geometry of
pregnant woman and children is also supplied, Fig. 7.1(a). The skeletal geometry and
the organ geometry of children is not available, however.
Fig. 7.1(b) shows the external and skeletal geometry of the 5th percentile female
(Robina). Fig. 7.1(c) shows the external and skeletal geometry of a six year old child
(Bobby). The external geometry (skin) was hand-scaled from an available eight year
old, while the skeleton was hand-scaled from the adult geometry to the anthropometric
dimension of the six year old child, where different body parts were scaled individually.
F IG . 7.1. Different HARB models (male, female, children). (Inset (a): Reproduced by permission of Digi-
mation/Viewpoint, http://www.digimation.com.)
F IG . 7.2. HARB models in car environments (Robby and H-models). (Inset (d) right-hand side: Reproduced
by permission of Prof. Kim, Kwangwon University.)
Crash test simulations with HARB models. Fig. 7.2 presents the different HARB mod-
els (Robby family) in a passenger car environment, with and without airbags in driver
and passenger positions, including a three year old in passenger airbag out-of-position
(OOP) posture (H-Model family).
Fig. 7.3(a) shows displaced shape snapshots of a frontal impact sled test simulation
for the seat belted 50th percentile male (Robby) with a deploying driver airbag.
F IG . 7.3. HARB models in frontal crash test situations (Robby and H-Models).
Fig. 7.3(b) repeats this scenario for the belted 5th percentile female (ROBINA [1998])
who is not protected by an airbag. Fig. 7.3(c) shows the six year old HARB model
(Bobby) in an OOP posture on the passenger seat. If the passenger side airbag fires at
that moment, the subject is hit by the bag during its deployment phase, which can cause
“bag slap” injuries and overextension and torsion of the neck.
Although the six year old should normally not be in the front seat, unbelted and
leaning forward, car industry is interested in studying such events in an effort to render
the airbag deployment phase as little aggressive to out-of-position subjects as possible,
while keeping the airbag effective under standard conditions. Since the standard me-
chanical dummies are not biofidelic for such events, simulation with human models is
recognized as a welcome alternative. Fig. 7.3(d), finally, shows a three year old HARB
model (H-Model family) undergoing the same passenger airbag OOP scenario.
Validation of HARB models. Fig. 7.4 summarizes validations performed on the exam-
ple of the model of the 50th percentile Robby model of the HARB family (ROBBY 1
[1997], ROBBY 2 [1998]). Validation of the other models were performed in a similar
way.
Validation of biomechanical simulation techniques (algorithms, codes, models) in
general can best be performed against physical test results. Directly exploitable phys-
ical results of crash events involving humans are rarely available. Access to physical
results can be gained through accidentological studies, accident reconstruction analy-
ses, cadaver test simulations and, rarely, volunteer test results for low energy impacts
(rear impact).
The physical validation test results of Fig. 7.4 were obtained from an existing
data base of cadaver tests performed at the University of Heidelberg (e.g., S CHMIDT,
K ALLIERIS , BARZ , M ATTERN , S CHULZ and S CHÜLER [1978]). The subject selected
from this data base was close to a 50th percentile male. The figure compares global
head, thorax and pelvis accelerations with the corresponding cadaver test results.
In view of the large scatter expected from different human subjects, the comparison
is considered satisfactory. In fact, validations should be made against experimental test
corridors obtained from a sufficiently large subset of near 50th percentile human sub-
jects. If the numerical simulation result curves remain within the given test corridors,
the validation is successful. Moreover, the numerical results should not be performed on
one “average” model of a given percentile slot, but simulations with stochastic variants
of the average model should match the physical test corridors. This requirement is ab-
sent when working with mechanical dummies, because mechanical dummies of a given
percentile category are all alike.
References on multi-body and muscle models. The following references may be con-
sulted on the subjects discussed in this chapter: BAUDRIT, H AMON , S ONG , ROBIN
and L E C OZ [1999] on dummy vs. human comparison; B EAUGONIN , H AUG and
H YNCIK [1998] on the Robby family of articulated human models; C HOI and L EE
[1999b] on a deformable human model; C OHEN [1987] on the analysis of frontal im-
pacts; D ENG [1985], D ENG and G OLDSMITH [1987] on the head/neck/upper torso
response to dynamic loading; D IGIMATION /V IEWPOINT C ATALOG [2002] contains
CAD data of the human body; G RAY ’ S A NATOMY [1989]; H AUG , L ASRY, G ROE -
NENBOOM , M UNCK , ROGER , S CHLOSSER and RÜCKERT [1993], H AUG [1995] and
H AUG , B EAUGONIN , T RAMECON and H YNCIK [1999] on biomechanical models for
vehicle accident simulation; H ILL [1970] on muscle tests; H UANG, K ING and C A -
VANAUGH [1994a], H UANG , K ING and C AVANAUGH [1994b] on human models for
284
E. Haug et al.
C HAPTER II
F IG . 7.4. 50th percentile HARB model validation with Heidelberg sled test results (Robby model).
side impact; H YNCIK [1997], H YNCIK [1999a], H YNCIK [1999b], H YNCIK [2000],
H YNCIK [2001a], H YNCIK [2001b], H YNCIK [2002a], H YNCIK [2002b], H YNCIK
[2002c] on human articulated rigid body models and deformable abdominal organ
models; I RWIN and M ERTZ [1997] on child dummies; K ISIELEWICZ and A NDOH
[1994] on critical issues in biomechanical tests and simulation; K ROONENBERG , A.
VAN DEN , T HUNNISSEN and W ISMANS [1997] on rear impact human models; L IZEE
and S ONG [1998], L IZEE , ROBIN , S ONG , B ERTHOLON , L E C OZ , B ESNAULT and
L AVASTE [1998] on deformable human thorax and full body models; M A , O BERGE -
FELL and R IZER [1995] on human joint models; M AUREL [1998] on a model of the
human upper limb; ROBBINS [1983], ROBBINS , S CHNEIDER , S NYDER , P FLUG and
H AFFNER [1983] contains basic human percentile data; ROBBY 1 [1997], and ROBBY 2
[1998] on the 50th percentile Robby multi-body models; ROBIN [2001] on the HUMOS
model; ROBINA [1998] on the 5th percentile female multi-body model of the Robby
family; S CHMIDT, K ALLIERIS , BARZ , M ATTERN , S CHULZ and S CHÜLER [1978] on
loadability limits of human vehicle occupant; S EIREG and A RVIKAR [1989] on muscle
models; S TÜRTZ [1980] on biomechanical data of children; V ISIBLE H UMAN P ROJECT
[1994] is a CD ROM with anatomy data of the human body; W ILLINGER , K ANG and
D IAW [1999] on 3D human model validation; W INTERS and S TARK [1988] (muscle
properties); W ITTEK and K AJZER [1995], W ITTEK and K AJZER [1997], W ITTEK ,
H AUG and K AJZER [1999], W ITTEK , K AJZER and H AUG [1999], W ITTEK , O NO and
K AJZER [1999], W ITTEK , O NO , K AJZER , Ö RTENGREN and I NAMI [2001] on active
and passive muscle models.
C HAPTER III
Deformable Human Models

8. The HUMOS human models for safety
8.1. Introduction
International projects, like HUMOS-1, pave the way for standardizing the modeling
methodology of the human body for impact biomechanics and occupant safety. The
major foreground information produced by the first HUMOS project is therefore not
specific computer models, but a database for the modeling of the human body with
commercially available dynamic crash software. The encryption of the contents of the
data base into specific versions of the individual commercial crash codes is a secondary
result of the project. If models made with commercial software make use of the data
published in the HUMOS database, if their modeling techniques are conform to the
rules fixed in the base, and if they pass the standard tests laid down in this base, then
the models could obtain the “HUMOS label”. This label would certify the model to be
conform to the standards laid down in the data base.
The HUMOS project embodies the fact that multi-disciplinary and multi-national
efforts are needed to eventually achieve the goal of the industrial use of internationally
standardized human models for safe car design. Putting humans into crash tests is only
feasible in the computer. The methodology of “virtual testing” using compute models
and codes, therefore, must be “certified” when it should complement or replace certain
legal crash dummy tests. There is at present no officially binding certification procedure
for simulation, neither for mechanical dummy nor for human models. The HUMOS
projects can help to establish such procedures.
The following paragraphs are extracted from, and prepared after, reports of the
HUMOS-1 project and the paper written by the HUMOS-1 project coordinator:
Robin, S. (2001), HUMOS: Human Model For Safety – A Joint Effort Towards the
Development of Refined Human-Like Car Occupant Models, 17th ESV Conference,
Paper Number 297.
The selected paragraphs and material intend to briefly summarize the contents and
major results of the first HUMOS project. The extracted material, figures and tables are
reproduced by permission of the ESV Conference Organizers.
In the first European HUMOS program (1999–2001) the development of commonly
accepted human car occupant models and computer methods was attacked. It was clear
from the outset, that the results of the first HUMOS program must be extended into
287
288 E. Haug et al. C HAPTER III
a second program, where other percentile human models and models with different
human morphologies should be created by appropriate modeling, scaling and morphing
techniques, and where the material description base should be enlarged.
Fourteen partners were involved, including six car manufacturers and several suppli-
ers, software houses, public research organizations and universities. The program dealt
with the synthesis and further development of the current knowledge of the human body
in terms of geometry, cinematic behavior, injury threshold and risk, the implementation
of this knowledge in new human body models, the development of the utilities for the
design office use, and delivering of the models to be available for their integration in the
car design process.
A wide bibliographical review supported those major goals. The geometry acquisi-
tion of a mid-sized 50th percentile male in a car driver seated position was achieved. The
main human body structures were then reconstructed using a CAD method. The mesh-
ing of the different structures was based on the CAD definition and has led to models
accounting for skin, bones, muscles as well as the main organs (lungs, heart, liver, kid-
neys, intestine etc.). The validation process was undertaken on a segment basis, each
main part of the human body being confronted with the available literature results. The
assembly of the whole model was the conclusion of this program.
The first problem that had to be solved was the acquisition of the inner and outer
geometry of a human, seated in a car occupant posture (the driver posture was chosen).
The publication made by ROBBINS , S CHNEIDER , S NYDER , P FLUG and H AFFNER
[1983] on the seated posture of vehicle occupants served as a basis for many devel-
opments of human substitutes. This work mainly qualified the position of external
anatomical landmarks on many different car occupants. But there were some limita-
tions in this work, particularly concerning the relative position of the different bony
structures and of the different organs, that had to be explored by the project.
8.2. Geometry acquisition
Selection of the subject. The geometry acquisition of a seated near 50th percentile
adult male is one of the original achievements of the HUMOS-1 program. Commonly
available human geometrical databases (e.g., GEBOD) furnish the main external dimen-
sions of the human being. Little information is available concerning the geometry of the
different organs and their relative positions of the different structures in a seated posi-
tion. Table 8.1 summarizes the main characteristics of the selected HUMOS subject and
those of the 50th percentile European male.
Subject frozen in driving position and physical slicing. The method consisted of plac-
ing the chosen cadaver in a driving position, Fig. 8.1.
Fig. 8.2 gives an overview on the applied methodology. The selected subject was in-
stalled in a standard full-size car cockpit with the hands placed on the steering wheel,
and it was frozen in this position, Fig. 8.2, inset(1). A reference frame related to the
cockpit was defined and the subject was embedded in a polymer block, Fig. 8.2(2). The
method of physical slicing of a frozen cadaver was chosen, Fig. 8.2(3). Each slice was
5 mm thick, and the saw was 2.5 mm thick (268 slices were made). The different slices
S ECTION 8 Deformable Human Models 289
TABLE 8.1
Main characteristics of the HUMOS subject and 50th percentile adult male (ROBIN
[2001])
Sitting height [mm] Standing height [mm] Weight [kg]
HUMOS subject 920 1730 80

50th percentile 915 1750 75.5
European male
F IG . 8.1. HUMOS subject in its driving position. (Reproduced by permission of ESV Conference Orga-
nizers.)
were photographed on each side (491 images) and then each image was hand-contoured
by anatomists, Fig. 8.2(4). This acquisition phase generated a set of about 13.000 files
representing 300 different organs. Each file was composed of a set of points character-
ized by their 2D coordinates related to the slice, for example, Fig. 8.2(5). The position
of each slice with regard to the reference frame was also available in the different files.
A 3D visualization, Fig. 8.2(6), based on the nodes, was carried out in order to validate
the acquisition process and, if need be, to modify some of the points describing the
organs.
3D geometrical CAD reconstruction process. Results of the 3D CAD reconstruction

process are shown on Fig. 8.3. During this process the point-by-point description files
were transformed into CAD geometrical files. Standard commercially available CAD
software was used. The main organs were reconstructed using available mathematical
surface definitions. A back and forth process was set up between anatomists and CAD
engineers in order to double-check at each step the shape of the reconstructed surfaces.
When some assumptions had to be made during this phase, they were thoroughly dis-
cussed with the anatomists.
At this point it should be kept in mind that human occupant models for safety should
be “generic”, rather than “specific” in the sense that “average” people should be placed
in cars during crash simulations. In reality, only specific subjects are available for data
acquisition, with particular defaults and deviations from average. It is therefore legiti-
mate to correct the acquired data for obvious deviations from the presumed anatomical
F IG . 8.2. Overview on the HUMOS-1 geometry acquisition process (HUMOS-1 project).
F IG . 8.3. Geometry acquisition result of (a) the whole body (without skin), (b) the skeleton and (c) the major
organs. (Reproduced by permission of ESV Conference Organizers.)
standard. In fact, human models for safety should be as standardized as possible. Defin-
itions of standard human anatomies have yet to be made. These goals must be achieved
in follow-up projects.
8.3. Meshing process
During meshing, a close collaboration was maintained with the anatomists in order to
double-check the shape of the reconstructed organs as well as the validity of the organ
connectivity assumptions. Only half of the skeleton was meshed and sagittal symme-
try was assumed. Further assumptions were made for some parts of the human body,
which were not described accurately enough. For example, the posterior arcs of the dif-
ferent vertebrae were not always visible, and not all the ligaments and muscles were
digitized during the acquisition process. Thus, some assumptions based on anatomical
descriptions were made in order to add some muscle parts and some ligaments.
For reasons of computational efficiency it was decided to try to keep the number of
deformable finite elements below 50 000, and requirements were issued on the final time
step (cf. Appendix A) of the complete model, which put a lower limit on the size of the
elements. Fig. 8.4 illustrates different segment meshes developed during this process.
F IG . 8.4. Mesh of skeleton and soft tissue parts. (Reproduced by permission of ESV Conference Organizers.)
It should be noted that the resulting mesh is relatively rough, and that finer meshes of
parts might be substituted in the future for refined studies.
The bony part of the skull was modeled with standard thin shells. The cervical spine
was modeled with shells and solids, some muscles and ligaments with bars. The remain-
ing spinal column has rigid vertebrae and nonlinear joint elements for the inter-vertebral
discs.
The final mesh of the whole thorax is a good example for the chosen mesh quality.
The clavicle, sternum, ribs and connective cartilaginous structures were modeled using
standard shell and solid finite elements. The scapula and the different vertebrae were
modeled using shell elements for the definition of the shape of those structures. In the
mechanical model, these bony parts are described as rigid elements. The muscular struc-
tures were meshed using both bar and solid elements. The flesh and connective tissues
were modeled using solid elements.
The other body parts were meshed with comparable quality.
8.4. Material laws
A literature review enabled to gather the available knowledge, cf. Appendix B. Due to
missing information, a limited number of experimental static and dynamic tests were
performed within the program.
Ribs, clavicles, sternum and pelvic bone were tested extensively in order to gain
specific knowledge about their mechanical properties. The results were associated with
the definition of a parameterized material law (a power law was used).
Furthermore, static and dynamic muscle compression experiments were carried out.
The strain-rate sensitivity was quantified for the soft tissues. Existing material laws were
adapted and implemented in the different participating crash codes.
For the different organs, very few experimental results are available and there is an
obvious lack of knowledge in this field. Available linear visco-elastic laws were chosen
according to other modeling publications. Soft tissue characterization is still regarded
as a largely uncovered area.
8.5. Segment validation process
Single rib model validation. The material properties were mainly derived from a liter-
ature review. The response corridors published by L IZEE and S ONG [1998] and L IZEE ,
ROBIN , S ONG , B ERTHOLON , L E C OZ , B ESNAULT and L AVASTE [1998] were used for
validation. Some material tests were carried out during the HUMOS-1 project. As an
example, the rib models were first checked against static and dynamic tests carried out
on isolated human ribs.
The ribs were modeled using shell elements for the cortical bone and solid elements
for the trabecular bone. The thickness of the cortical bone was modeled using different
shell thickness depending upon the location. In order to account for rib fracture, a failure
plastic strain was introduced in the rib material law. The different values used for the
material behavior definition of the ribs are reported in Table 8.2.
TABLE 8.2
Material properties of the rib cortical bones (ROBIN [2001])
Rib cortical bone material model: elastic-plastic behaviour with failure
Elasticity modulus 14 GPa

Yield Stress 70 Mpa
Ultimate stress 70 Mpa
Maximum deformation 4%
Poisson’s ratio 0.3
Density 6000 kg/m3
F IG . 8.5. Computer result compared with experimental corridor of an isolated rib subjected to a quasi-static
loading. (Reproduced by permission of ESV Conference Organizers.)
These values were used for both the static and dynamic validation of the ribs. Fig. 8.5
illustrates the static behavior of the so-called “HUMOS rib” compared with the results
obtained on an isolated human rib by the University of Heidelberg.
Thorax model validation. For the bones and cartilaginous parts of the thorax for which
experimental data were available, the same process was used for the first description of
the model. The assembled segment models were validated against the published experi-
mental results. Figs. 8.6 and 8.7 represent the results obtained for the HUMOS-1 thorax
model (ESI Software version) as compared to the experimental pendulum impact test
results published by K ROELL , S CHNEIDER and NAHUM [1971], K ROELL , S CHNEI -
DER and NAHUM [1974] for frontal impact and by V IANO [1989] for lateral impact.
In these experiments, the thorax was impacted with a 23.4 kg cylindrical impactor at
4.9 m/s and 4.6 m/s, in frontal and lateral directions, respectively.
The organs of the thorax were modeled with the airbag modeling technique (lungs)
or as “bio-bags” (heart), cf. Appendix D. A bio-bag model is an adaptation of a standard
airbag model, for which the mechanical properties of the enclosed gas are modified to
be close to incompressible, like a fluid. This feature permits to approximately simulate
F IG . 8.6. HUMOS-1 FE model: frontal impact test (Kroell test) (HUMOS-1 project/ESI).
F IG . 8.7. Lateral impact test (Viano test) (HUMOS-1 project/ESI).

the impact response of hollow internal organs, filled partly or fully with air and liquid
contents. It also allows organ contents to escape (vent out) under compression of the or-
gan, when the standard venting options of the used airbag models are invoked. Standard
airbag models are supported by all dynamic crash codes. They model a compressible
gas contained in a flexible and extensible hull, made of membrane, shell or solid finite
elements. More elaborate modeling of the gas or fluid contents of the hollow organs will
require the techniques of fluid or particle dynamics, not discussed here.
The importance of at least approximate modeling of the blood contents of the heart is
demonstrated in Fig. 8.6, where the impact force versus sternum deflection curves are
shown both, for an empty and for a full heart.
Abdomen model validation. The test reported by C AVANAUGH , N YQUIST, G OLD -

BERG and K ING [1986] was simulated with the HUMOS model (ESI Software version),
Fig. 8.8. In this test, a bar (∅25 mm, length 381 mm) with total mass of 32 kg hits the
abdomen with a velocity of 6.9 m/s. The simulation outcome depends largely on the
way the abdominal cavity and the intestines are modeled, Fig. 8.8(a).
The abdominal cavity is treated like an equivalent bio-bag, closed by the abdominal
walls, and filled with a fluid and with sub-models of the intestines and internal organs.
The intestines can either be modeled as solids, Fig. 8.8(a)(1), as an inner simple (2) or
“meandered” (3) bio-bag, each placed within the outer bio-bag of the abdominal cavity.
F IG . 8.8. Abdomen bar impact test (Cavanaugh test) (HUMOS-1 project/ESI).

The solid model of the intestines gave too much shear and compressive stiffness at
large deformations, Fig. 8.8(c). The model with the bio-bag responded well initially, but
developed too high internal fluid pressures at large deformations. The intestine model
with a meandered bio-bag gave the best results and the response curve falls in between
the experimental test corridor. The results show that even for a faithful reproduction
of gross results, such as external load-deformation curves, the models should follow as
closely as possible the anatomical realities, which allow for major shifts and relative
displacements of the inner organs.
8.6. Discussion and conclusions
The HUMOS-1 program led to a first definition of a finite element model of the human
body in a seated driving posture. In an effort to create a European standard human
model, useful for car in industry, the model was implemented with three commercially
available dynamic crash codes (Madymo, Pam–Crash and Radioss). The mesh of the
model is shared by the different software packages, but the validation of the different
models was carried out separately for the different codes.
From the beginning of this research work, it was foreseen that some major limita-
tions would be encountered. First, the geometrical definition of the model comes from
a unique human subject, i.e., is not generic. It is expected in a follow-up program to
develop scaling techniques, which would enable to first define a generic 50th percentile
model from the current reference mesh, and second 5th and 95th percentile occupant
models. It is also expected to be able to derive from this first model some pedestrian
models.
A validation database was built and used in order to validate the different segments of
the model. The global validation of the whole model remains to be done. Some investi-
gations need to be carried out on the muscle tonus contribution, especially for the low
speed impact conditions that can be encountered in real field accident analysis. Further-
more, some limitations are due to the lack of knowledge of the injury mechanisms. The
main currently used criteria were implemented in the model, but its injury prediction
capabilities are limited with regard to its limited complexity.
8.7. Acknowledgements
The HUMOS project was funded by the European Commission under the Industrial and
Materials Technologies program (Brite–EuRam III). The consortium partners were the
LAB (Laboratory of Accidentology and Biomechanics PSA Peugeot, Citroën, Renault),
who coordinated the work and was involved mainly in the meshing process of the tho-
rax and in the validation database. The other participating car manufacturers were Volvo
(meshing of the neck), BMW (meshing of the upper limbs), and VW (literature review
of the existing models). Software developers were strongly involved in this program.
ISAM/MECALOG (Radioss software) carried out the 3D CAD reconstruction of the
model, the head mesh description, and the assembly of the final model with Radioss.
ESI Software (Pam–Crash software) modeled the lower limbs and was in charge of the
homogeneity of the different segment models. ESI Software also performed the assem-
bly of the final model with Pam–Crash. TNO (Madymo Software) carried out modeling
work and coordinated the soft tissue behavior activities. The supplier FAURECIA car-
ried out the pelvis and abdomen modeling and coordinated the geometry acquisition
process. INRETS LBMC (Laboratoire de Biomécanique et Mécanique des Chocs) car-
ried out full-scale sled tests with human substitutes and contributed to the extension of
the validation database. Marseille University was in charge of the geometrical acqui-
sition of the seated human body, carried out by INRETS LBA (Laboratory of Applied
Biomechanics). Athens University defined the physical material laws for the different
human soft tissues and carried out experiments on some muscle properties. Heidel-
berg University was in charge of experimental investigations on different human tissues
(bones and some cartilage structures) and produced a set of new experiments for the
extension of the validation database. Chalmers University performed a bibliographical
study on the current knowledge about human tissue behavior and identified the fields of
missing knowledge. Chalmers University also contributed in the validation of the neck
model.
8.8. References for the HUMOS model
The following references were found essential for the work reported in this section:
C AVANAUGH , N YQUIST, G OLDBERG and K ING [1986] on the impact response and
tolerance of the human lower abdomen; K ROELL , S CHNEIDER and NAHUM [1971],
K ROELL , S CHNEIDER and NAHUM [1974] on the impact response and tolerance of
the human thorax; L IZEE and S ONG [1998], L IZEE , ROBIN , S ONG , B ERTHOLON ,
L E C OZ , B ESNAULT and L AVASTE [1998] on 3D FE models of the thorax and whole
body; M ERTZ [1984] on a procedure for normalizing impact response data; ROBBINS ,
S CHNEIDER , S NYDER , P FLUG and H AFFNER [1983] on the seated posture of vehicle
occupants; ROBIN [2001] on the HUMOS model for safety; V IANO [1989] on biome-
chanical responses and injuries in blunt lateral impact.
9. The fiftieth percentile male H-Model
9.1. Introduction
Preamble. This section gives an overview on the structure of the fiftieth percentile
male H-Model of the human body. This family of models is presently distributed and
developed by the private companies ESI Software (Paris) and IPS International (Seoul),
and by Hong-Ik University (Seoul). The presently available basic model represents a
50th percentile male human body and it was conceived primarily to study injury mech-
anisms and to assess injuries of the human skeleton and organs, which result from car
accidents, including pedestrian injuries, e.g., C HOI and L EE [1999b]. A fifth percentile
female and a 3 and 6 year old child model are under development. The basic model
permits a fast and ongoing absorption of the rapidly growing biomechanical research
results. The model can assist the conception of improved crash protection measures and
devices for the human driver and passenger. It allows the omni-directional analysis for
different impact directions (front, side, rollover, etc.) with a single model with global
and local responses to bags, belts, head restraints, knee bolsters, etc. The analysis of
different body postures, and the effect of muscle activity an be studied using the model.
From the H-Models other models for the analysis of passenger sitting comfort and
riding comfort can be derived. The models for sitting comfort permit the assessment
of seat and backrest pressure and they are whole body models equipped with elaborate
representations of the flesh. The models for riding comfort permit to assess the effect of
seat vibrations and they have well adapted spines and organ masses to detect resonance.
An attempt is made to systematically present the material for each body part in para-
graphs which discuss anatomy, injury, model structure and calibration, and validation.
The discussed items are the whole body H-ARB models and models of the head, neck,
torso, upper extremity, lower extremity and the foot/ankle complex.
Features. Major features of the H-Model are given by the possibility of modular
assembly of the deformable external and internal components into an underlying multi-
body H-ARB (Human Articulated Rigid Body) model. The H-ARB model assures a
correct overall kinematic behavior in a simulated crash scenario. Modularity permits
to study each body part of interest in detail. While assuring the good overall kinemat-
ics, detailed studies can so be made for the head in head impacts, the neck in whiplash
events, the thorax and abdomen for belt injuries, the lower extremity in knee bolster
impacts, the foot/ankle complex in toe panel intrusions and the upper extremity and
shoulder complex in side impacts and airbag aggressions. Fig. 9.1 shows an overall
view of the model with skin and skeleton.
Model structure. The H-Model is a recent combination of two previous biomechanical

models of the human body, namely the “ROBBY” (ESI Software) and the “H-Dummy”
(IPSI/Hong-Ik) model families. The overall anthropometric properties of both mod-
els were mainly extracted from the report edited by ROBBINS [1983], after which
the ROBBY family had been named. The new model family is now co-developed by
F IG . 9.1. H-Model with skin and skeleton (fiftieth percentile male).

the Seoul and Paris biomechanics teams. The development of the H-Model family was
started prior to and is independent of the HUMOS-1 model described in Section 8 of
this article.
The geometry of all listed models and sub-models has been derived mostly from
the anatomy data found in the CAD geometry data base issued by DIGIMATION
(D IGIMATION /V IEWPOINT C ATALOG [2002]) and in the CD ROM issued by the V IS -
IBLE H UMAN P ROJECT [1994]. Further inspiration, concerning mainly the anatomical
details of the ligaments, tendons, muscles, blood vessels and other soft tissues and inter-
nal organs was found in anatomical atlases, such as G RAY ’ S A NATOMY [1989]. Where
necessary, the data were scaled to the presumed dimensions of the 50th percentile male.
The Viewpoint CAD data were transformed into surface finite element meshes, and
three-dimensional bulk FE meshes were generated in a compatible fashion, using com-
mercially available FE mesh generator packages. The Visible Human CT data were
digitized manually. The digitized anatomical cross sections were used to generate sur-
face and bulk FE meshes.
The HUMOS-1 project (cf. Section 8) used a specific slicing technique, applied on a
PMHS (post mortem human subject) in a sitting driver position. This procedure enabled
the project to produce unique data of a seated subject. Since only one subject could be
treated, scaling to the fiftieth percentile was necessary.
Calibration. All listed models and sub-models have been given material properties
found in the open literature. A description of bio-materials is given in Section 1.5 and
in Appendix B of this article.
Validation. All listed models and sub-models have been validated with available litera-
ture results. The basic validation tests and results are described in the available H-Model
Reference Manuals. All validation test cases are relatively simple, representative, re-
peatable and well documented and controlled experiments, which facilitates modeling
under well identified conditions.
9.2. The H-ARB Human Articulated Rigid Body model
Model structure and calibration. The H-ARB versions of the H-Models consist of ar-
ticulated rigid body segments with flexible joints. The fiftieth percentile male model,
H-ARB50, was built after ROBBINS [1983] and is documented in the H-Model Ref-
erence Manuals. It represents the basic platform for the modular assembly of detailed
deformable and frangible skeleton components with soft tissues and internal organs. It
serves for the evaluation of the overall kinematic and kinetic behavior of the occupant
for omni-directional impacts. Fig. 9.2 shows the H-ARB50 version of the H-Model
(76 kg) with the joint tripods (a) indicating the location and the orientations of the
anatomical joints.
The outer geometry of the skin is taken from the clay model published in ROBBINS
[1983], Fig. 9.2(b). The major dimensions of the outer surface of the model are shown
in inset (c). The neck of the H-ARB model has several joints (skull-C1, C1-C2, etc.)
as shown in inset (c), but no muscles. The various joints have rotational moment/angle
F IG . 9.2. H-ARB50 version of the H-Model (fiftieth percentile male). (Inset (b): from ROBBINS [1983]
(UMTRI-83-53-2, US DoT NHTSA public domain report).)
relationships indicated in ROBBINS [1983] (see also Appendix B). Two typical curves
are shown in Fig. 9.3. Some joints are given damping coefficients in order to stabilize
their dynamic response. The joints are modeled with standard joint finite elements with
linear and nonlinear moment-angle curves.
Accelerometers. Three accelerometers, which can be used to evaluate global motions

of the H-ARB models, are defined using three additional nodal points. The locations of
the accelerometers is adopted in analogy to the Hybrid III 50% male dummy model.
F IG . 9.3. H-Model H-ARB50 joint moment-angle curves.
The solver code will specifically output the acceleration time histories at these precise
locations.
Contact. Contact surfaces in are defined between different segments of the H-Model,
such as arms to thorax, between upper legs, etc. Note that in any application prob-
lem, additional contact interfaces involving the H-Models should be defined, such as
H-Model to airbag, to belt, to car seat, etc. Standard contact algorithms are used, as
implemented in the commercial crash codes (cf. Appendix A, Section A.3).
Validation and performance. First, the extension motion of the neck due to rear impact
between two cars is presented. The geometry of the animated neck motion is shown in
Fig. 9.4 at different times. The acceleration pulses of the striking and the hit car, the
input acceleration pulse applied at T1 (first thorax vertebra), and the linear acceleration
responses of the human volunteer and of the H-ARB50 head model are also shown in
the figure. Any active muscle action is not considered in this simple H-ARB model. The
passive muscle action, and the action of the discs and neck ligaments, is considered in a
global fashion via the modeled standard joint finite elements with calibrated linear and
nonlinear moment-angle curves.
Second, for the lateral bending motion of the neck due to the side impact, the ani-
mated neck motion for different times is shown in Fig. 9.5. The input acceleration pulse
given at T1 (first thorax vertebra) is shown in the figure. The responses of the H-ARB50
neck model is compared with corridors of human volunteers.
Third, for a frontal sled test simulation, Fig. 9.6 represents the overall response of the
H-ARB50 model to the medium intensity acceleration diagram indicated in inset (a).
The H-ARB response is compared at different times to the response of a Hybrid III 50th
percentile dummy model, inset (b). The head and upper torso linear acceleration versus
time diagrams of both articulated rigid body models are compared in inset (c), where the
differences in the responses reflect the differences between humans and crash dummies.
Finally, for an arbitrary passenger out-of-position (OOP) scenario, Fig. 9.7 shows
the overall response of the H-ARB version of the H-Model.
F IG . 9.4. H-ARB50 Validation: Horizontal acceleration histories due to rear impact.
Zooms with FE sub-models. For detailed studies (“zooms”) of the impact response of
the individual body parts, it is possible to substitute their detailed finite element mod-
els into the H-ARB versions of H-Model. The following sections outline such detailed
sub-models. This modular approach saves CPU time while the overall model response
correctly represents the kinematics of the body to generate the dynamic boundary con-
ditions for the required zoom.
References on multi-body human models. To obtain further information on the shown

HARB models, the following references may be consulted (in addition to the ref-
erences listed at the end of Chapter II in the paragraph “References on multi-body
and muscle models”): JAGER , S AUREN , T HUNNISSEN and W ISMANS [1994], JAGER
[1996] on head/neck models; P RASAD [1990] on comparison between Hybrid dummy
performances; SAE E NGINEERING AID 23 [1986] is the Hybrid III User’s Man-
ual; T HUNNISSEN , W ISMANS , E WING and T HOMAS [1995] on human volunteers in
whiplash loading; Z INK [1997] on 6 year old out-of-position passenger airbag simula-
tion.
9.3. H-Head: Skull and brain
An overview on the H-Head model structure and application is given in C HOI [2001].
F IG . 9.5. H-ARB50 Validation: Neck motion due to side impact.
Anatomy. The complex anatomy of the skull and brain is summarized in Fig. 9.8.
More details can be found in anatomical atlases and textbooks, e.g., G RAY ’ S A NATOMY
[1989] and in the V ISIBLE H UMAN P ROJECT [1994]. The finite element model of the
human head (H-Head) has been constructed by evaluating cross sections through the
head. Three such sections (top-down) and the corresponding traces made of the brain
matter are shown in Fig. 9.8(a) (S PITZER and W HITLOCK [1998]).
Cranial structure. The skull consists of three layers referred to as the outer table,
diploe, and inner or vitreous table, Fig. 9.8(b) (P IKE [1990]) and (e) (P UTZ and PABST
[2000]). The diploe consists of trabecular bone, and is located between the other two
that are made up of compact (cortical) bone. The inside surface of the cranial cavity
is lined by a layer of dense fibrous irregular connective tissue and enclosing venous
sinuses. This layer adheres for the most part tightly to the cranial bones.
F IG . 9.6. H-ARB50 Validation: Sled test for frontal impact.
F IG . 9.7. OOP passenger airbag with H-ARB50.

F IG . 9.8. Anatomy of the human head (overview). (Inset (a): Reproduced by permission of Jones & Bartlett
Publishers Inc.; Inset (b): Reprinted with permission from “Automotove Safety” ©1990 SAE International;
Insets (c) and (e): Reproduced by permission of Urban & Fischer Verlag; Inset (d): Reproduced by permission
of Wally Welker at The Department of Neurophysiology, The University of Wisconsin; Inset (f): Graphic by
John Sundsten, courtesy of the Structural Informatics Group at the University of Washington.)
Intra-cranial structure. The intra-cranial structure contains mainly the brain, the
cerebrospinal fluid (CSF), the vascular structure, and the membranous coverings (dura
mater, pia mater) and partition structures, Fig. 9.8(c)–(f) (P UTZ and PABST [2000] and
web sites http://brainmuseum.org/specimens/primates/human/brain/human8sect6.jpg
and http://www9.biostr.washington.edu/cgi-bin/DA/imageform). The brain occupies
F IG . 9.9. Brain injuries (http://www.trauma.org/imagebank/imagebank.html). (Reproduced by permission of

Trauma.Org.)
most of the cranium and is composed of right and left hemispheres, separated by a
fold of the dura mater called the falx cerebri. Another fold of the dura mater, called the
tentorium cerebelli, separates the cerebrum and the cerebellum. The junction between
the folds of the dura and the inner surface of the cranium forms some of the venous
sinuses. These sinuses receive blood drained from the brain and reabsorb cerebrospinal
fluid in regions knows as arachnoid granulations.
The next layers below the dura mater are the avascular arachnoid membrane and the
delicate pia mater. The space between the arachnoid and the pia mater is called the
subarachnoidal space, which is traversed by the arteries of the brain and cranial nerves
and contains cerebrospinal fluid. The CSF circulates around the brain and spinal cord
and through four ventricles (two lateral, the third ventricle at the mid-line below the
lateral ventricles, and the fourth ventricle between the brain stem and the cerebellum).
Head injury. Head injuries can be caused by external or internal loads. External loads
result from impacts with objects or obstacles, while internal dynamic loads result from
motions of the rest of the body and are transmitted to the head through the neck (e.g.,
whiplash). Major head injuries are considered skull fractures (bone damage) and brain
injuries (neural damage).
Skull fractures with increasing severity can be linear, comminuted (fragmented), de-
pressed and basal fractures. Skull fractures correlate weakly with neural damage, i.e.,
neural damage may exist with and without skull fracture.
Neural damage can result from diffuse and focal brain injuries, Fig. 9.9. Diffuse brain
injuries consist in brain swelling, concussion and diffuse axonal injury (DAI), while
focal brain injuries lead to subdural hematomas (SDH), epidural hematomas (EDH),
intra-cerebral hematomas (ICH), and contusions (from coup and countercoup pressure).
For example, upon high angular head acceleration associated with an impact, the bridg-
ing veins at the top of the brain can rupture, causing a subdural hematoma or blood clot
to develop. The clot exerts large pressures on the brain and if not quickly relieved can de-
prive the brain of its blood supply, causing brain death. The stretch in some of the bridg-
ing veins can reach 200%! (according to Viano: “humans are made to fall from trees”).
Model structure and calibration. The detailed FE head model (H-Head) of the 50th
percentile H-Model consists of deformable bone and brain components, Fig. 9.10. The
modeling of the head comprises the scalp (can attenuate blunt impacts), the skull and
F IG . 9.10. H-Head model overview.
facial bone (the latter can be held more approximate), the cerebro-spinal fluid (CSF)
layers and the ventricles, the brain membranes (dura mater), the falx, the tentorium, the
gray and the white brain matter and the cerebellum. The CSF is presently modeled with
the Murnaghan equation of state, p = p0 + B((ρ/ρ0 )γ − 1), where p is the hydrostatic
pressure, ρ is the mass density and B and γ are material constants. For the modeling
of the vacuum cavity that might develop in the CSF at the interface between the skull
TABLE 9.1
Mechanical properties of the H-Head model
Component E K G ν ρ
[kPa] [kPa] [kPa] [kg/m3 ]
Skull Outer table 7.3 × 106 0.22 3000

Inner table 7.3 × 106 0.22 3000
Diploe 2.02 × 105 1.39 × 105 0.22 1410
Facial bone 7.3 × 106 0.22 2700
Mandible 7.3 × 106 0.22 2700
Dura mater 3.15 × 104 0.22 1133
Venous sinus 1.0 × 105 1000
CSF 1.0 × 105 1000
Falx 3.15 × 104 0.45 1133
Pia 3.15 × 104 0.45 1133
Tentorium 3.15 × 104 0.45 1133
Brain Gray matter 7.96 × 103 7.96 × 101 ∼ 0.499 1040
White matter 1.27 × 104 1.27 × 102 ∼ 0.499 1040
Ventricle 1.0 × 105 1000
Cerebellum and stem 1.27 × 104 1.27 × 102 ∼ 0.499 1040
E = Young’s modulus [kPa], K = Bulk modulus [kPa], G = Shear modulus [kPa], ν = Poisson’s ratio, ρ =
Mass density [kg/m3 ].
and brain from negative countercoup pressures, ideal gas equations are used. The model
is fully compatible with the H-Neck spinal cord model. Table 9.1 shows the material
coefficients as calibrated for the H-Head model.
Validation. Fig. 9.11 summarizes a validation of the H-Head model carried out after
tests performed by Nahum, using frontal pendulum impact. In Fig. 9.11(a) the test setup
and the measured pressure time histories in coup, countercoup, parietal and occipital
locations are shown. Fig. 9.11(b) contains the measured and calculated head acceler-
ations. Fig. 9.11(c) gives an overview on the surface pressures as calculated with the
H-Head model at different times. In Fig. 9.11(d)–(f) the pressure time histories in coup,
countercoup and occipital locations are shown and comparison to Nahum’s tests is seen
to be satisfactory.
References on the H-Head model. The following references were considered relevant
for the basic and detailed understanding, construction and validation of the head model:
A BEL , G ENNARELLI and S EGAWA [1978] on incidence and severity of cerebral con-
cussion in rhesus monkeys from sagittal acceleration; BANDAK and E PPINGER [1994],
BANDAK [1996] on brain FE models and impact traumatic brain injury; C HAPON ,
V ERRIEST, D EDOYAN , T RAUCHESSEC and A RTRU [1983] on brain vulnerability from
real accidents; C HOI and L EE [1999b] on deformable FE models of the human body;
C LAESSENS [1997], C LAESSENS , S AUREN and W ISMANS [1997] on FE modelling
of the human head under impact conditions; C OOPER [1982a], C OOPER [1982b] on
injury of the skull, brain and cerebro-spinal fluid related; D I M ASI , M ARCUS and
E PPINGER [1991] on a 3D anatomical brain model for automobile crash loading;
F IG . 9.11. H-Head model validations (tests: after N AHUM , S MITH and WARD [1977]).
D ONNELLY and M EDIGE [1997] on shear properties of human brain tissues; E W-

ING , T HOMAS , L USTICK , M UZZY III, W ILLEMS and M AJEWSKI [1978] on the
effect of the initial position on the head/neck response in sled tests; G ENNARELLI
[1980], G ENNARELLI , T HIBAULT, A DAMS , G RAHAM , T HOMPSON and M ARCINCIN
[1982] on the analysis of head injury severity by AIS-80 (1980) and on diffuse ax-
onal injury and traumatic coma in primates (1982); G RAY ’ S A NATOMY [1989]: Atlas
of Anatomy; G URDJIAN and L ISSNER [1944], G URDJIAN , W EBSTER and L ISSNER
[1955], G URDJIAN , ROBERTS and T HOMAS [1966] on mechanisms of head injury
and brain concussion and tolerance of acceleration and intra-cranial pressure; H OL -
BOURN [1943] on the mechanics of head injury; K ANG , W ILLINGER , D IAW and
C HINN [1997] on the validation of a 3D FE model of the human head in motorcy-
cle accidents; L EE , M ELVIN and U ENO [1987], L EE and H AUT [1989] on FE analysis
of subdural hematoma and bridging vein failure characteristics; L ISSNER , L EBOW and
E VANS [1960] on experiments on intra-cranial pressure; M ARGULIES , T HIBAULT and

G ENNARELLI [1990] on modeling brain injury in primates; M ILLER and C HINZEI
[1997] on constitutive modelling of brain tissue; NAHUM , S MITH and WARD [1977]
on intra-cranial pressure dynamics during head impact; N EWMAN [1993] on head pro-
tection; O MMAYA , H IRSCH , F LAMM and M AHONE [1966], O MMAYA and H IRSCH
[1971], O MMAYA and G ENNARELLI [1974] on cerebral concussions in the monkey
(1966), on their tolerances (1971) and on their clinical/experimental correlation (1974);
O NO , K IKUCHI , NAKAMURA , KOBAYASHI and NAKAMURA [1980], O NO [1999] on
head injury tolerance for sagittal impact from tests (1980) and on spine deformation and
on vertebral motion from whiplash test volunteers (1999); P ENN and C LASEN [1982]
on traumatic brain swelling and edema; P UTZ and PABST [2000]: Sobotta Atlas of Hu-
man Anatomy; RUAN , K HALIL and K ING [1991] on human FE head model response
in side impacts; RUAN and P RASAD [1994] on head injury assessment in frontal im-
pacts by mathematical modelling; S ANCES ET AL . [1982] on head and spine injuries;
S COTT [1981] on the epidemiology of motor cyclist head and neck trauma; S PITZER
and W HITLOCK [1998]: Atlas of the Visible Human Male; TARRIERE [1981] on in-
vestigation of the brain with CT-scanners; T ORG [1982], T ORG and PAVLOV [1991] on
athletic injury on the head, neck and face; T URQUIER , K ANG , T ROSSEILLE , W ILL -
INGER , L AVASTE , TARRIERE and D ÖMONT [1996] on the validation of a 3D FE head
model against experiments; U ENO , M ELVIN , L UNDQUIST and L EE [1989] on 2D FE
analysis of human brain under impact; VOO , K UMARESAN , P INTAR , YOGANANDAN
and S ANCES [1996] on a finite element model of the human head; WALKE , KOLLROS
and C ASE [1944] on the physiological basis of concussion; W ILLINGER , K ANG and
D IAW [1999] on the validation of a 3D FE human model against experimental impacts;
W ISMANS ET AL . [1994] on injury biomechanics; Z EIDLER , S TÜRTZ , B URG and R AU
[1981] on injury mechanisms in head-on collisions; Z HOU , K HALIL and K ING [1996]
on the visco-elastic brain FE modeling for sagittal and lateral rotation acceleration. The
following sites were used: http://www9.biostr.washington.edu/cgi-bin/DA/imageform;
http://brainmuseum.org/specimens/primates/human/brain/human8sect6.jpg; http://www.
trauma.org/imagebank/imagebank.html.
9.4. H-Neck: Cervical spine with active muscles
Overviews on the H-Neck model structure and applications are given in C HOI and E OM
[1998], L EE and C HOI [2000] and C HOI , L EE and H AUG [2001b]. A study on the
whiplash injury due to the low velocity rear-end collision is presented in C HOI , L EE ,
E OM and L EE [1999].
Anatomy. The anatomy of the cervical spine is summarized in Fig. 9.12 (after
http://www.rad.washington.edu/RadAnat/Cspine.html and P UTZ and PABST [2000]).
The human cervical spine is composed of 7 vertebrae, C1–C7, with five similar lower
vertebrae (C3–C7) and two dissimilar upper vertebrae (C1 “atlas” and C2 “axis”). Ap-
proximately 47% of the bending and stretching of the cervical spine occurs between
the head and C1, while over 50% of the rotational motions occur between C1 and C2.
Between each pair of adjacent vertebrae except for between C1 and C2 exist a structure
known as the disc of cervical spine. Each disc is composed of the nucleus pulposus,
F IG . 9.12. Anatomy of the human cervical spine. (Inset (a) left: Reproduced by permission of Michael
L. Richardson, University of Washington, Deparment of Radiology; Insets (a) center, (b) and (c) left: Repro-
duced by permission of Urban & Fischer Verlag.)
the annulus fibrosus and a cartilageous end-plate. Between 70 to 90% of the nucleus
pulposus by weight is water, and takes up as much as 40 to 60% of the disc area.
The annulus fibrosus is a laminated and hence anisotropic structure composed of sev-
eral layers in which each layer maintains a 30◦ angle of inclination from the horizontal
plane. The inner boundary of the annulus fibrosus is attached to the cartilageous end-
plate, and the outer surface is directly connected to the vertebra body. The discs play a
dominant role under compressive loads.
The discs show greater stiffness for the front/rearward motion than for the side/side
motion. Spinal ligaments (anterior, posterior, . . .) connect the vertebrae of the cervical
spine. They ensure that the spinal motions occur within physiological bounds, thus pre-
venting spinal code injury due to excessive spinal motion. The ligaments are usually
situated between two adjacent vertebrae, and in some cases connects several vertebrae.
Neck injury. Neck injury from car accidents has been studied extensively and is sum-
marized hereafter according to M C E LHANEY and M YERS [1993] (eds. Nahum and
Melvin). Fig. 9.13, inset (a), summarizes the anatomical head motions. Inset (b) shows
the modes of loading on the neck. Inset (c) lists 10 distinct injury mechanisms and
identifies a total of 25 different neck injuries. Inset (d) shows three tension-extension
injury mechanisms, including the important whiplash injury (B). Inset (e) shows three
flexion-compression neck injury mechanisms.
For example, the tension–extension injury (A) of inset (d) to Fig. 9.13 is described
as “Fixation of the head with continued forward displacement of the body. This occurs
commonly in unbelted occupants hitting the windshield, and as a result of falls and
dives”. The tension–extension whiplash injury (B) is described as “Inertial loading of
the neck following an abrupt forward acceleration of the torso as would occur in a
rear-end collision”. The tension–extension injury (C) is described as “Forceful loading
below the chin directed postero-superiorly (as in a judicial hanging).
The frequent whiplash injury occurs as a result of more or less mild rear-end colli-
sions. It is considered a hyperextension injury and it may cause muscle stiffness and
neck pain. Larger accelerations, in addition to producing whiplash symptoms, may
produce disruption of the anterior longitudinal ligament and intervertebral disk, and
horizontal fractures though the vertebra. The remaining neck injuries are described in
the above reference. The survivable neck injuries from car accidents are often injuries
of the soft connecting tissues, while bone fractures are less common.
Model structure and calibration. The modeling aspects of the human cervical spine
are summarized in Figs. 9.14 and 9.15. The H-Neck model is fully compatible and inter-
faced with the H-Head model. Based on 3D CAD data of 50% male cervical vertebrae
supplied by DIGIMATION (D IGIMATION /V IEWPOINT C ATALOG [2002]), the cervical
spine in the current H-Neck model is built as a chain of articulated rigid vertebrae. This
reduces computational cost, all the while soft tissue injury plays a more important role
than bony fracture does in neck injury. Fully deformable vertebra bodies can easily be
modeled, for example, as demonstrated earlier by N ITSCHE , H AUG and K ISIELEWICZ
[1996].
The H-Neck model contains important elements such as vertebrae, discs, ligaments,
and stabilizing muscles. The inter-vertebral contacts are modeled with sliding inter-
faces and contact stiffnesses. They can also be modeled by appropriate layers of finite
elements. The vertebrae of the H-Neck model are interconnected by the neck ligaments,
F IG . 9.13. Neck injury overview (after M C E LHANEY and M YERS [1993]). (Reproduced by permission
of Springer Verlag from: Nahum and Melvin eds., Accidental Injury, Biomechanics and Prevention: Insets
(a) and (b): Figure 14.5, page 318; Inset (c): Table 14.5, page 319; Inset (d): Figure 14.10, page 324; Inset (e):
Figure 14.8, page 320.)
F IG . 9.14. (C3–C4) disc 6 DOF kinematic joint element and ligament calibration (H-Neck).
modeled as bars or membranes. To account for the visco-elastic and anisotropic bio-
mechanical properties of the ligaments, nonlinear one-dimensional tension-only bar
elements are used. For cruciform, transverse, and tectorial ligaments in the atlanto-
axial joint however, membrane elements are used to allow for surface contact (see also
Fig. 9.12(c)). Each pair of adjacent vertebrae (except C1 and C2) is connected by the
elastic discs of the cervical spine.
In the H-Neck model, the complex deformation of the discs is modeled by 6-DOF
nonlinear spring elements with stiffness calibrated as shown in Fig. 9.14(a), (b). The
ligament force-displacement calibrations are shown in Fig. 9.14(c).
These calibrations were made with test results found in the literature, for example:
C HAZAL , TANGUY and B OURGES [1985] on biomechanical properties of the spinal
ligaments; M C C LURE , S IEGLER and N OBILINI [1998] on in vivo 3D flexibility prop-
erties of the human cervical spine; M ORONEY, S CHULTZ , M ILLER and A NDERSSON
[1988], M ORONEY, S CHULTZ and M ILLER [1988] on lower cervical spine properties
and neck loads; M YKLEBUST and P INTAR [1988] on the tensile strength of spinal lig-
aments; N IGHTINGALE , W INKELSTEIN , K NAUB , R ICHARDSON , L UCK and M YERS
[2002] on upper/lower cervical spine strength comparison in flexion/extension; PAN -
JABI , C RISCO , VASAVADA , O DA , C HOLEWICKI , N IBU and S HIN [2001] on cervical
spine load-displacement curves; YOGANANDAN , S RIRANGAM and P INTAR [2001] on
the biomechanics of the cervical spine.
Fig. 9.15(a) shows the anatomical details (P UTZ and PABST [2000]), the H-Neck
ligaments and the assembled H-Head/H-Neck model, where the head can be modeled
effectively as a rigid body for most whiplash simulations. The muscle forces obey the
well known active and passive Hill muscle law, Fig. 9.15(b). The muscles are modeled
by nonlinear bar elements, which can be curved in space, assuring a correct introduction
F IG . 9.15. H-Neck model overview. (Inset (a) left: Reproduced by permission of Urban & Fischer Verlag.)
of the muscle forces along the curved neck line. The active muscle force can contribute
to conscious bracing and motion. The passive muscle forces contribute to the resistance
of the connected body parts near the end of their respective anatomical motion ranges.
It increases with muscle length in a highly nonlinear fashion (cf. Appendix C).
Validation. Fig. 9.16 shows in inset (a) the flexion response of the H-Neck model in a
frontal passenger car impact simulation. In inset (b) the extension response in whiplash
for a rear impact is shown. Inset (c) shows pictures of the deformed neck in lateral
bending as from a side impact. The muscle bars are seen to follow the neck as it flexes
and extends.
Fig. 9.17 contains several results of the H-Head/H-Neck response to a frontal car
crash event, as compared to the response corridors obtained with volunteers in live tests
(after T HUNNISSEN , W ISMANS , E WING and T HOMAS [1995] and H ORST, T HUNNIS -
SEN , H APPEE , H AASTER and W ISMANS [1997]).
F IG . 9.16. H-Neck model validations: Flexion, extension and torsion motions.
Fig. 9.18 shows the trajectories of several points of the H-Neck model in extension
from rear impacts (whiplash). The results compare well with corridors from tests per-
formed with life volunteers (O NO , K ANEOKA , W ITTEK and K AJZER [1997]).
Fig. 9.19(a)–(c) compare the responses of the H-Neck model mounted on the H-
ARB version in a rear impact sled test simulation, both, with and without a headrest.
Fig. 9.19(d) compares the H-Head/H-Neck response to a rear impact with a headrest in a
low and a high position. Fig. 9.20, finally, shows the head and the cervical spine skeleton
response with a deformable headrest to such a rear impact. The results demonstrate the
utility of such models to study the effects of preventive measures in the design of passive
safety restraints of passenger cars.
JARI volunteer tests. Recently the H-Neck™ model is being validated against volun-
teer neck whiplash tests, Fig. 9.21, in which human volunteers underwent equivalent
mild frontal and rear head accelerations in the sagittal plane, as they are typical in mild
frontal and rear impacts of passenger cars, O NO , K ANEOKA , S UN , TAKHOUNTS and
E PPINGER [2001]. While the evaluations of the tests are still ongoing, some simulations
of the scenarios have been made with an attempt to study the influence of voluntary and
involuntary contraction of the neck muscles which counteract the accidental neck/head
motions in the sagittal plane, Fig. 9.21(a). The neck muscles are divided into groups of
neck flexors and neck extensors in the sagittal plane, insets (b) and (c), which are acti-
vated separately in a parametric study in different assumed reflex time scenarios to their
F IG . 9.17. H-Neck model validations: Volunteer tests in frontal crash (flexion). (Test results after Stapp
papers T HUNNISSEN , W ISMANS , E WING and T HOMAS [1995] and H ORST, T HUNNISSEN , H APPEE ,
H AASTER and W ISMANS [1997].)
highest level of activation (100%). The flexor group pulls the head forward (agonists)
and the extensor group backward (antagonists).
For example, the subject’s head is being suddenly pulled rearward with the help of
a string attached to a mask, which contains accelerometers and optical gauges, as shown
F IG . 9.18. H-Neck model validations: Trajectories in extension for a rear impact (whiplash). (Right inset:
Reproduced by permission of The Stapp Association.)
in inset (a). The volunteer is either asked to relax, or to fully stiffen its neck muscles
prior to the sudden pull. An intermediate case is when the subject involuntarily stiffens
its neck muscles during the event (stretch reflex). In inset (d) deformed shape snapshots
of a simulation of a relaxed subject with no activation of the active muscle forces are
shown. In (e) the neck flexor muscles are assumed to be fully tensed, starting at 100 ms
after the pull, followed by the extensor muscles, starting at 150 ms. As expected, the
maximal neck flexion occurs for the relaxed subject, (d), with trajectories, (f), and this
motion is reduced for the subject with the (arbitrarily) assumed pattern of full muscle
activation, (g), (h). Since the exact pattern of involuntary muscle activation is unknown,
such activation patterns can be found by back calculation from various assumed simple
activation patterns and comparison with volunteer tests. Pattern (i) was not a good as-
sumption, since the late activation of the flexors allowed for large neck flexion motions.
The active muscle forces were built up and applied according to Hill’s law, using the
physiological cross section area of each muscle. Although it is presently unknown in
precisely what pattern a living subject activates the neck muscles in a whiplash event,
such parametric studies can help clarify the open questions. Light can also be shed on
the different loads and injuries of the cervical spine, which are influenced by the muscle
actions.
References on the H-Neck model. The following references were considered useful
for the preparation of the H-Neck model: B OKDUK and YOGANANDAN [2001] on mi-
nor injuries of the cervical spine; C USICK and YOGANANDAN [2002] on major injuries
of the cervical spine; D OHERTY, E SSES and H EGGENESS [1992] on odontoid frac-
ture; DVORAK , H AYEK and Z EHNUDER [1987], DVORAK , PANJABI and F ROEHLICH
[1988] on rotary instability of the upper cervical spine; F OREMAN and C ROFT [1995]
(whiplash injuries); G RAY ’ S A NATOMY [1989]: Atlas of Anatomy; H IRSCH [1955]
(disc compression reaction); H ORST, T HUNNISSEN , H APPEE , H AASTER and W IS -
MANS [1997] on the influence of muscle activity on head/neck impact response;
H UELKE , N USHOLTZ and K AIKER [1986] on an overview on cervical fracture and dis-
location; JAGER [1996] on the mathematical modeling of the cervical spine (overview);
F IG . 9.19. H-Neck model validations: Low and high headrest scenarios.
JAGER , S AUREN , T HUNNISSEN and W ISMANS [1994] on head/neck models; K U -

LAK , B ELYTSCHKO , S CHULTZ and G ALANTE [1976] on non-linear behaviour of discs
under axial load; K ALLIERIS and S CHMIDT [1990] on side impact neck response
from cadaver tests; M ARKOLF and M ORRIS [1974] on components of intervertebral
discs; M C E LHANEY, PAVER , M C C RACKIN and M AXWELL [1983] on cervical spine
compression response; M YERS , M C E LHANEY, R ICHARDSON , N IGHTINGALE and
D OHERTY [1991], M YERS , M C E LHANEY, D OHERTY, PAVER and G RAY [1991] on
the influence of end conditions (a) and torsion (b) on cervical spine injury; N OVOTNY
F IG . 9.20. H-Neck model validations: Rear impact whiplash response with deformable headrest.
[1993] on spinal biomechanics; O NO [1999] on spine deformation and vertebral motion

from whiplash test volunteers; PANJABI , DVORAK and D URANCEAU [1988] on 3D
movements of the upper cervical spine; P ENNING [1979], P ENNING and W ILMARK
[1987] on movements and rotations of the cervical spine; P UTZ and PABST [2000]:
Sobotta Atlas of Human Anatomy; S ONNERUP [1972] on cervical disc analysis in com-
pression; S VENSSON and L ÖVSUND [1992] on new dummy neck for rear end collision;
T ENNYSON and K ING [1976a], T ENNYSON and K ING [1976b] on biodynamic model
of the spinal column and measure of muscle action; T HUNNISSEN , W ISMANS , E WING
and T HOMAS [1995] on human volunteers in whiplash loading; V IRGIN and L UDHI -
ANA [1951] on physical properties of intervertebral discs; W HITE and PANJABI [1990]
on clinical biomechanics of the spine; W INTERS and S TARK [1988] on muscle prop-
erties; W ISMANS , VAN O ORSCHOT and W OLTRING [1986] on omnidirectional human
head-neck response; W ITTEK and K AJZER [1995], W ITTEK and K AJZER [1997], W IT-
TEK , H AUG and K AJZER [1999], W ITTEK , K AJZER and H AUG [1999], W ITTEK , O NO
and K AJZER [1999], W ITTEK , O NO , K AJZER , Ö RTENGREN and I NAMI [2001] on ac-
tive and passive muscle models; YOGANANDAN , S ANCES and P INTAR [1989b] on axial
compression of human cadaver and manikin necks.
The following web sites were consulted: http://www.rad.washington.edu/RadAnat/
Cspine.html.
9.5. H-Torso: Rib cage, spine and thoracic organs
An overview on the H-Torso model is given in C HOI and L EE [1999a].
Anatomy of the spine and torso. The anatomy of the human spine and thorax
is summarized in Fig. 9.22(a)–(f), prepared with the web sites http://www9.biostr.
washington.edu/cgi-bin/DA/imageform, (a) and (b), and http://www.nlm.nih.gov/
research/visible/image/abdomen_mri.jpg, (c), of the National Library of Medicine
(NLM).
The human spine consists in the flexible cervical spine (neck; seven vertebrae), the
more rigid thoracic spine (twelve vertebrae with attached ribs), the flexible lumbar spine
F IG . 9.21. Volunteer test simulation for sudden rearward pulling on subjects head mask (H-Neck).
F IG . 9.22. Anatomy of the human torso. (Insets (a) right and (b): Reproduced by permission of the University
of Washington Digital Anatomist; Inset (c): National Library of Medicine (NLM) public domain information.)
(five vertebrae), the rigid sacrum (five fused vertebrae) and the coccyx (tail bone; four
fused segments of bone). The spine supports the upper body, it protects the spinal chord
and it provides mobility via its twenty four movable and nine fixed vertebrae. The ver-
tebrae are hinged together by their facet joints and by the inter-vertebral discs. The facet
joints, discs and ligaments of the spine limit its overall deformation in bending, torsion
and elongation.
The human chest mainly consists of the ribs, the thoracic spine and the cardio-
pulmonary organs. The upper ten ribs are directly or indirectly connected to the sternum
via their costal cartilages, while the lowest two ribs are floating. The twelve ribs of
each side articulate posterior with the twelve thoracic vertebrae. Three layers of in-
tercostal muscles connect the ribs. The mediastinum forms the space between the
cardio-pulmonary organs, consisting of trachea, lungs, aorta and the heart. The esoph-
agus transverses the thorax. The diaphragm separates the thorax region from the upper
abdomen. The thorax serves as a support for the shoulders and the upper extremities.
The abdominal cavity, Fig. 9.22(c)–(f), is separated from the thorax by the diaphragm.
The abdomen can be subdivided into the upper and lower abdomen. It contains the
“solid” organs: liver, spleen, pancreas, kidneys, adrenal glands and ovaries; and the
“hollow” organs: stomach, small and large intestines, urinary bladder and the uterus
(ROUHANA [1993], H YNCIK [1999b], H YNCIK [2001b], H YNCIK [2002b], H YNCIK
[2002c]). These organs are less well protected for impact than the organs of the thorax.
Injury to the thoraco-lumbar spine. According to K ING [1993], injury to the bony part
of the thoraco-lumbar spine is rare in frontal car accidents. Injury can be classified into
anterior wedge fractures and burst fractures of vertebral bodies, dislocations and frac-
F IG . 9.23. Chest and upper abdomen injury. (Insets (a) and (f): Reproduced by permission of Trauma.Org;
Inset (b): Reproduced by permission of the Yale University School of Medicine; Inset (c): C AVANAUGH
[1993], reproduced by permission of Springer Verlag from: Nahum and Melvin eds., Accidental Injury, Bio-
mechanics and Prevention, Figure 15.4, page 367; Insets (d) and (e): from CIREN William Lehman Injury
Research Center at the Ryder Trauma Center Case # 97-018 (public domain material, U.S. Department of
Transportation, NHTSA), see NHTSA/CIREN [1997].)
ture dislocations, rotational injuries, Chance fractures, hyper-extension injuries and soft
tissue injuries. A typical vertebra fracture is shown in the inset to Fig. 9.23(a). Anterior
wedge fractures occur under combined axial compression and bending (pilot ejection)
and consist in a lighter injury. Burst fractures threaten the integrity of the cord and occur
at higher loads. Dislocations can produce misalignments of the facets between adjacent
vertebrae and occur under combined flexion, rotation and postero-anterior shear and
can involve bone fracture and ligament rupture, threatening the cord. Rotational injuries
come from excessive torsion of the spine under simultaneous axial and shear loads, with
lateral wedge fractures and a danger of paraplegia. Chance fractures (C HANCE [1948])
occur by over-flexion of the lumbar spine when wearing only a lap belt in a frontal col-
lision and it splits a lumbar vertebra in its transverse plane. If the upper body is properly
fixed, excessive flexion over the restraining lap belt cannot occur. Hyper-extension in-
juries occur in pilot ejection and rarely in car accidents. They may produce bone fracture
and ligament rupture. Soft tissue injuries involve the intervertebral discs, the ligaments,
the joint capsule, the facet joints and the muscles and tendons attached to the spinal
column.
Thorax/abdominal injury. For occupants of passenger cars, most crash injuries of the
thorax are given by single and multiple rib fracture (“flail chest”), which result from
the compression of the rib cage and the ensuing bending and breaking of the ribs,
Fig. 9.23(a) (http://www.trauma.org). Flail chest prevents the rib cage from acting as
a coherent protective cage and impairs respiration. Broken ribs can cause injury to
the organs inside the rib cage, either from unrestrained large dynamic compression,
or from punctures caused by the sharp ends of broken ribs. For the thorax region,
pulmonary artery and lung injuries, and liver and heart injuries are frequently encoun-
tered. Lung contusions and laceration of lung tissue due to broken ribs can lead to
hemothorax (bleeding of the lung tissue) and to pneumothorax (pneumatic collapse of
one or both lungs) from punctures of the pleural sac between the lungs and the rib
cage, Fig. 9.23(b), respectively (http://info.med.yale.edu/intmed/cardio/imaging/cases/
pneumothorax_tension/graphics/rad1.gif).
The heart, Fig. 9.23(c) (from C AVANAUGH [1993]), letter A, can suffer contusion
from compression and high rates of compression, and high compression of the sternum
can lead to laceration. Rupture of the aorta, B, occurs frequently at its root, near the
ligamentum arteriosum (isthmus), C, and at its insertion into the diaphragm, E. In lat-
eral impacts, aorta rupture between its unsupported upper and its more firmly anchored
descending branch, D, can result from lateral motions of the heart. The aorta walls act
anisotropic in failure, with more frequent failure in transverse direction.
Trauma of the organs of the abdomen, Figs. 9.23(d) and (e) (http://www-nrd.nhtsa.dot.
gov/departments/nrd-50/ciren/ciren1.html) and (f) (http://www.trauma.org/~imagebank/
chest/images/chest0023c.jpg), may stem from penetrating objects, or blunt impacts, the
latter being predominant in transport vehicle accidents, including injuries from the lap
belt. The abdominal cavity is filled by the solid organs (liver, spleen, pancreas, kidneys,
adrenal glands, ovaries) and by the hollow organs (stomach, large and small intestines,
urinary bladder, uterus). These organs are only lightly tethered.
They can slide easily within their membrane envelopes (peritoneum) and they are held
in place by their blood supply vessels (kidneys), or by peritoneal folds and ligaments.
The large relative mobility of the abdominal organs makes injury sensitive to the attitude
of the subject at the time of the accident. Blunt trauma affects the solid organs more than
the hollow organs. Side impacts favor renal injury. The action of lap belts frequently
leads to minor injuries, while preventing from serious injuries.
Model structure and calibration. The modeling aspects of the human torso are sum-
marized in Figs. 9.24 and 9.25.
The thoracic and lumbar spine model is shown in Fig. 9.24.
The thorax of the H-Model mainly consists in deformable and damageable rib and
rigid spine bones and in the thoracic organs and great vessels with internal air (lungs) or
blood modeling, Fig. 9.25(a) (http://www.bartleby.com/107/illus490.html) and (b). The
blood-filled vessels are simulated with an incompressible fluid model, while the lungs
can compress like airbags. Both models can leak out fluid or compressible air (“bio-
bags”, cf. Appendix D). If impacts on the thorax are simulated, the abdominal cavity
can be represented approximately as an incompressible bio-bag (see Fig. 9.27(b), (d)).
In Section 8.5 on the HUMOS model validation, Fig. 8.6(c) demonstrates the effect of
F IG . 9.24. H-Spine model: Thoracic and lumbar spine. (Insets (c) and (d): Anatomy after P UTZ and PABST
[2000]: Reproduced by permission of Urban & Fischer Verlag.)
simulating the blood fill of the heart on the Kroell frontal pendulum impact test. If the
heart was modeled as an empty bag with elastic skin, the resulting chest deformation
was over-predicted.
Fig. 9.25(c) depicts the finite element models of the upper abdominal organs of the
H-Model. The hollow organs (stomach, etc.) are modeled with bio-bags, while the solid
organs (liver, etc.) are modeled with solid finite elements. An ad hoc modeling of the
viscera was introduced in Section 8.5, Fig. 8.8(a), (d). Further investigation will clarify
to which detail the visceral organs must be modeled for various purposes. If the up-
per abdominal organs are studied, the viscera can often be modeled approximately as
bio-bags. Similar models were established in collaboration with the University of West
Bohemia, Prof. Rosenberg, by H YNCIK [1999b], H YNCIK [2001b], H YNCIK [2002b],
H YNCIK [2002c].
Fig. 9.26 summarizes the modeling and calibration aspects of the ribs of the H-Torso
model. Fig. 9.26(a) shows the more complex thin shell solid models and the simpler
tapered beam model of the ribs. The rib bone material properties, including for rib
fracture, were calibrated against test results, Fig. 9.26(b) and (c).
F IG . 9.25. H-Torso model (thorax and upper abdomen). (Inset (a) anatomical drawing: Reproduced by per-
mission of Bartleby.com, Inc.)
Validation. For a frontal pendulum impact loading, Fig. 9.27 shows the response
of the H-Torso model after the cadaver (and animal) tests done by K ROELL [1971],
K ROELL , S CHNEIDER and NAHUM [1971], K ROELL , S CHNEIDER and NAHUM
[1974], K ROELL , A LLEN , WARNER and P ERL [1986]. Fig. 9.27(a) represents the test
setup and the resulting pendulum force-chest deflection corridor. Fig. 9.27(b)–(d) dis-
play the H-Torso model response to the applied frontal pendulum impact loading. In
(c) and (d) the deformations of the heart, lungs and blood vessels are clearly visible.
Fig. 9.27(e) shows the response of the H-Torso rib cage with von Mises surface stress
contours added on the ribs and with the failure of some ribs (“flail chest”). Fig. 9.27(f),
finally, shows the H-Torso model pendulum force versus thorax deflection curve to fall
well into the response corridor obtained from many cadaver tests by Kroell. Similar
good responses (not shown here) were obtained for lateral pendulum impacts (Viano
tests), similar as shown in Section 8.5, Fig. 8.7, for the HUMOS model segment valida-
tion.
For chest belt loading, Fig. 9.28 summarizes the response obtained with the H-Torso
model.
Fig. 9.28(a) shows the corresponding cadaver test setup (C ESARI and B OUQUET
[1990], C ESARI and B OUQUET [1994]) with the measured cadaver and Hybrid III
dummy chest band and thorax deflection versus impact energy results. The chest band
device is an elastic steel band that is wrapped tightly around the chest. Strain gauges
measure the flexural deformation from which the deflected shape of the tape, and hence
the thorax, can be deduced. Fig. 9.28(b) shows the H-Torso model global deformation
response to the applied belt loading. The response of the ribs, including fracture near
maximum deformation, are shown in Fig. 9.28(c). Fig. 9.28(d) and (e) compare chest
band results from the cadaver tests with the response of the H-Torso model, where the
chest band results are approximated by plotting the external model section contours at
the locations of the upper and lower chest bands, respectively. The agreement is deemed
satisfactory, although the model should have duplicated the real chest bands with their
true physical properties and contact with the chest for a closer representation.
For abdomen bar impact tests, validations of the H-Model similar as the ones carried
out in Section 8.5 for the HUMOS model were performed (not repeated here).
References on the H-Torso model. The following references were consulted for the
establishment of the H-Torso model (first authors): A LLAIN [1998] on thorax model cal-
ibration; A LLEN , F ERGUSON , L EHMANN and O’B RIEN [1982] on fracture/dislocation
of the lower cervical spine; B EGEMAN , K ING and P RASAD [1973] on spinal loads from
front/rear acceleration; B ELYTSCHKO , K ULAK , S CHULTZ and G ALANTE [1972], B E -
F IG . 9.26. H-Torso rib modeling and calibration. (Inset (b) left diagram: Reprinted with permission from
SAE paper number 856027 © 1985 SAE International; Inset (c): X-ray picture: Reproduced by permission of
Trauma.Org.)
F IG . 9.27. H-Torso frontal chest pendulum impact validation. (Inset (a): Reproduced by permission of
The Stapp Association.)
LYTSCHKO , K ULAK , S CHULTZ and G ALANTE [1974] on FE analysis of intervertebral

discs; B OUQUET, R AMET, B ERMOND and C ESARI [1994] on thorax/pelvis response to
impact; C ESARI and B OUQUET [1990] and C ESARI and B OUQUET [1994] on thoracic
belt loading; C HAZAL , TANGUY and B OURGES [1985] on spinal ligament proper-
ties; C LEMENTE [1981] (anatomy atlas); C OOPER , P EARCE , S TAINER and M AYNARD
[1982] on thorax trauma and cardiac injury; E PPINGER [1976], E PPINGER , M ARCUS
and M ORGAN [1984] on thoracic injury and dummy development; F UNG and Y EN
[1984] on lung injury experiments; G OEL , G OYAL , C LARK , N ISHIYAMA and N YE
[1985] on lumbar spine kinematics; H UELKE , N USHOLTZ and K AIKER [1986] on
thoraco-abdominal biomechanics research; K APANDJI [1974c] (joint physiology of the
trunk and vertebral column); K AZARIAN , B EERS and H ERNANDEZ [1979], K AZAR -
IAN [1982] on the spine injuries; K EITHEL [1972] on thoraco-lumbar intervertebral
joint deformation due to loads; K LEINBERGER , S UN , E PPINGER , K UPPA and S AUL
[1998] on improved injury criteria; K ULAK , B ELYTSCHKO , S CHULTZ and G ALANTE
F IG . 9.28. H-Torso belt loading validation. (Inset (a), insets (d) and (e) test results: Reproduced by permis-
sion of The Stapp Association.)
[1976] on non-linear behaviour of discs under axial load; L ASKY, S IEGEL and NAHUM
[1968] on automotive cardio-thoracic injuries; M AKHSOUS , H ÖGFORS , S IEMIEN ’ SKI
and P ETERSON [1999] on shoulder strength in the scapular plane; M ARKOLF [1972],
M ARKOLF and M ORRIS [1974] on deformations under loads of thoraco-lumbar in-
tervertebral joints and disc structure; M ILLER , S CHULTZ , WARWICK and S PENCER
[1986] on lumbar spine properties under large loads; M OFFATT, A DVANI and L IN
[1971] on experiment and analysis of the human spine; M YKLEBUST and P INTAR
[1988] on the tensile strength of spinal ligaments; NAHUM , G ADD , S CHNEIDER and
K ROELL [1970], NAHUM , S CHNEIDER and K ROELL [1975] on the response of the
human thorax under blunt impact; N EUMANN , K ELLER , E KSTROM , P ERRY, H ANS -
SON and S PENGLER [1992] on the properties of the human lumbar anterior ligament;
N ICOLL [1949] on fracture of the torso lumbar spine; N OVOTNY [1993] on spinal bio-
mechanics; PANJABI and B RAND [1976] on the mechanical properties of the human
thoracic spine; P LANK , K LEINBERGER and E PPINGER [1994] on FE analysis of tho-

rax/restraint system interaction; P OPE , K ROELL , V IANO , WARNER and A LLEN [1979]
on the postural influence on thoracic impact; P RASAD and K ING [1994] on an exper-
imentally validated dynamic spine model; ROAF [1960] on the mechanics of spinal
injury; S CHNEIDER , K ING and B EEBE [1990] on thorax-abdomen trauma assessment;
S TALNAKER and M OHAN [1974] on human chest impact protection criteria; S TOCK -
IER , E PSTEIN and E PSTEIN [1969] on seat belt trauma to the lumbar spine; V ERRIEST
and C HAPON [1994] on thoracic injury criteria and rib fractures; V IANO and L AU
[1983], V IANO [1989] on the influence of impact velocity and chest compression in tho-
rax injury (1983) and on the response and injuries blunt lateral impact (1989); W HITE
and PANJABI [1978], W HITE and PANJABI [1990] on spine biomechanics; YOGANAN -
DAN , H AFFNER , M ALMAN , N ICHOLS , P INTAR , J ENTZEN , W EINSHEL , L ARSON and
S ANCES [1989], YOGANANDAN and P INTAR [1998] on the trauma of the human spine
(1989a) and on the biomechanics of thoracic ribs (1998).
The following web sites were consulted: http://info.med.yale.edu./intmed/cardio/
imaging/cases/pneumothorax_tension/graphics/rad1.gif; http://www.bartleby.com/107/
illus490.html; http://www.bionetmed.com/; http://www.nlm.nih.gov/research/visible/
image/abdomen_mri.jpg; http://www.trauma.org; http://www9.biostr.washington.edu/
cgi-bin/DA/imageform.
9.6. H-UE: Shoulder and arms
An overview on the human upper extremity model structure and application is given in
C HOI , L EE and H AUG [2001a].
Anatomy. The anatomy of the human upper extremity is summarized in Fig. 9.29.
The upper extremity consists of the arm and the shoulder complex, Fig. 9.29(a) (after
http://www.rad.washington.edu/RadAnat and P UTZ and PABST [2000]). The shoulder
(scapula, clavicula) is connected to the thorax via the scapula-thoracic (sliding) joint
and via the sterno-clavicular (ball) joint. The upper arm connects to the scapula by the
gleno-humeral joint. The scapula connects to the clavicula by the acromio-clavicular
joint.
The skeleton of the upper extremity is connected by several systems of mus-
cles, Fig. 9.29(b) (after http://www.fitstep.com/Advanced/Anatomy/Shoulders.htm) and
Fig. 9.29(c). These muscles connect the upper arm to the body (latissimus dorsi,
pectoralis major), the upper arm to the scapula (supraspinatus, infraspinatus, teres,
subscapularis, deltoid, biceps, triceps, coraco-brachialis, brachialis), the scapula to the
thorax (trapezius, levator anguli, rhomboids, serratus, pectoralis minor), the upper arm
to the clavicula (deltoid, pectoralis major) and the clavicula to the thorax (subclav-
ius, trapezius). The bones of the upper extremity are further connected by ligaments,
Fig. 9.29(d) (after http://eduserv.hscer.washington.edu/hubio553/atlas/shjointlig.html).
The ligaments that are connecting the clavicula to the ribs are not shown.
Injury. The injury of the upper extremity is discussed in part in L EVINE [1993]. The
upper extremity is most frequently hurt through side impacts between two vehicles,
F IG . 9.29. Anatomy of the human upper extremity. (Inset (a) X-ray pictures: Reproduced by permission
of Michael L. Richardson, University of Washington Medical Center, Department of Radiology; Inset (a)
anatomical drawing: Reproduced by permission of Urban & Fischer Verlag; Insets (b) and (c): Reproduced by
permission of BetterU, Inc.; Inset (d): Reproduced by permission of Carol C. Teitz, University of Washington.)
where the occupant may hit the door in a lateral motion. The forearm may be affected
when it is positioned over the steering wheel while the driver side airbag deploys. The
upper extremity may also be injured from violent projection by side impact airbags. Be-
cause they must fire faster, side impact airbags turn out to be more aggressive than front
impact airbags. Upon side impact, the clavicle with its acromio-clavicular and sterno-
clavicular joints may be injured through fracture and dislocations. When the lower arm
is placed over a driver side airbag deploying from the steering wheel, there may be frac-
tures of the arm. Upon inflation of side airbags located in the back rest, the upper arm
may experience fracture and the gleno-humeral joint may become dislocated. The in-
juries can be more or less severe, ranging from dislocations to closed fractures to severe
open fractures.
Model structure and calibration. The modeling aspects of the human upper extrem-
ity are summarized in Fig. 9.30. The upper extremity (H-UE) of the H-Model mainly
consists of deformable and damageable bones and flesh padding. The flesh padding
uses solid elements (not shown). Active and passive muscle forces are modeled with
bar elements using the Hill muscle model. Nonlinear contact interfaces model the car-
tilage layers on the shoulder and elbow joints. Fig. 9.30(a) shows the H-UE model in
its bodily context, while Fig. 9.30(b) shows the skeleton and the attached long muscle
bars. Fig. 9.30(c) contains the modeling of the major joint ligaments, which are modeled
with short nonlinear bars. Fig. 9.30(d) shows the modeling of surface muscles with two-
dimensional finite elements and superimposed muscle bars. Here membrane elements
are used as a “matrix” onto which a number of Hill type muscle bars are attached like
“fibers” in series and in parallel, as required by the surface curvature and the surface
area occupied by each muscle segment. The membranes represent the passive muscle
material properties perpendicular to the fibers, while the bars represent the active and
passive longitudinal muscle fiber properties. This way the curvature and the sliding
contact of the muscles with the rib cage and other parts of the skeleton is well modeled.
Fig. 9.30(e) shows the behavior of this model under imposed motions of the arm and
shoulder. The next step might be to adapt “composite” materials, with the transverse and
bulk properties of the muscles represented by the composite matrix, and the active and
passive muscle fiber properties assigned to the fiber phase of the composite materials.
Fig. 9.31 shows, for example, the arm bone material calibration from three point
bending tests on cadaver arms, after P INTAR , YOGANADAN and E PPINGER [1998]
(public domain). The tests, Fig. 9.31(a), were simulated up to fracture, Fig. 9.31(b), and
F IG . 9.30. H-UE modeling.
the calculated calibrated and the measured force-displacement curves are compared in
Fig. 9.31(c).
Validation. Based on PALANIAPPAN , W IPASURAMONTON , B EGEMAN , TANAVDE

and Z HU [1999], Fig. 9.32 shows the calibration of an earlier version of the model on
F IG . 9.31. H-UE arm bone material calibration.
a pendulum impact experiment. The hands were modeled rigid, the deformable bones
were modeled with shells and the flesh was modeled with deformable solid elements.
The flesh model was calibrated on the pendulum impact tests. The attenuation of the
shocks through the deformable flesh was found to be an important element to achieve
the required accuracy of the model.
F IG . 9.32. H-UE pendulum impact validation (PALANIAPPAN, W IPASURAMONTON, B EGEMAN, TANAVDE

and Z HU [1999]). (Reproduced by permission of The Stapp Association.)
From the same reference, Fig. 9.33 validates the H-UE model on a side impact airbag
out of position deployment. The tests were performed at Wayne State University. In
Fig. 9.33(a) the human arm model was mounted on a model of a fiftieth percentile
Hybrid III dummy model. The left arm of the driver is out-of-position and it is projected
forward by a deploying side impact airbag. This airbag is mounted in the backrest of
the driver seat. In inset (b) the calculated and the measured arm displacement responses
are compared.
References on the H-UE model. The following references were consulted when
establishing the H-UE model (first authors): E NGIN [1979], E NGIN [1980], E N -
GIN [1983], E NGIN [1984], E NGIN and P EINDL [1987], E NGIN and C HEN [1989],
E NGIN and T’ MER [1989] on works on the biomechanics of the human shoul-
F IG . 9.33. H-UE airbag impact validation (PALANIAPPAN , W IPASURAMONTON , B EGEMAN , TANAVDE

and Z HU [1999]). (Reproduced by permission of The Stapp Association.)
der complex, bone and joint resistance; G OLDSTEIN , F RANKENBURG and K UHN
[1993] on the biomechanics of bone; I RWIN [1994] on shoulder and thorax re-
sponse cadaver tests and their analysis; K APANDJI [1974a] (upper limb joint phys-
iology); P EINDL and E NGIN [1987] on passive resistive properties of the shoulder
complex; P RADAS and C ALLEJA [1990] on the non-linear visco-elastic properties
of human hand flexor tendon; P UTZ and PABST [2000]: Sobotta Atlas of Hu-
man Anatomy; T’ MER and E NGIN [1989] on a 3D model of the shoulder com-
plex. The following web sites were used: http://www.rad.washington.edu/RadAnat;
http://www.fitstep.com/Advanced/Anatomy/Shoulders.htm; http://eduserv.hscer.
washington.edu/hubio553/atlas/shjointlig.html.
9.7. H-LE: Knee-thigh-hip complex
An overview on the H-LE model and applications is provided in YOO and C HOI [1999].
F IG . 9.34. Anatomy of the human lower extremity (without foot/ankle). (Insets (a) and (f) X-ray pictures:
Reproduced by permission of Michael L. Richardson, University of Washington Medical Center, Department
of Radiology; Insets (a) and (f), anatomical drawings, and insets (b) and (d): Reproduced by permission of
Urban & Fischer Verlag; Insets (c) and (e): Figs. 7.16a, p. 205, 7.16d, p. 206 and 8.14c, p. 231 from HUMAN
ANATOMY, 4th ed. by Frederic H. Martini, Michael J. Timmons and Robert H. Tallitsch, Copyright © 2003
by Frederic H. Martini, Inc. and Michael J. Timmons.)
Anatomy. The anatomy of the human lower extremity and pelvis is summarized in
Fig. 9.34. It consists of hip, upper and lower legs, ankles and foot, articulated by the
hip, knee, and ankle joints (the foot/ankle complex is discussed in a separate section).
The pelvis supports the spinal column and it contains the sacrum, the coccyx, and the
two hip bones, each made of the three fused ilium, ischium and pubic bones, Fig. 9.34(a)
(http://www.rad.washington.edu/RadAnat/pelvis.html and P UTZ and PABST [2000]),

Fig. 9.34(d) (P UTZ and PABST [2000]) and Fig. 9.34(e) (M ARTINI , T IMMONS and
TALLITSCH [2003]). These three pelvic bones converge at the acetabulum, the articula-
tion for the head of the femur, (a). The femur (thigh bone), (b) (after P UTZ and PABST
[2000]), is the longest bone in the body. Its lower end joins the tibia (shin) to form the
knee joint. Its upper end is rounded into a ball (or “head” of the femur) that fits into a
socket in the pelvis (the acetabulum) to form the hip joint, (d), (e).
The neck of the femur gives the hip joint a wide range of movement, but it is a point
of weakness and a common site of fracture.
The tibia is the inner and thicker of the two long bones in the lower leg. The tibia
runs parallel to the smaller and thinner fibula to which it is attached by ligaments. The
upper end of the tibia joins the femur to form the knee joint. The lower ends of tibia and
fibula form the ankle joint with the medial and lateral malleolus, Fig. 9.34(c) (M ARTINI ,
T IMMONS and TALLITSCH [2003]).
The knee joint, Fig. 9.34(f) (http://www.rad.washington.edu/RadAnat/knee.html and
P UTZ and PABST [2000]), is held together by flexible ligaments. The collateral liga-
ments run along the sides of the knee and limit sideways motion. The anterior cruciate
ligament (ACL) limits rotation and relative forward motion of the tibia. The posterior
cruciate ligament (PCL) limits relative backward motion of the tibia. The lateral menis-
cus and medial meniscus are pads of cartilage that cushion the joint, acting as shock
absorbers between the bones. The patella is the roughly triangular-shaped bone at the
front of the knee joint. It transmits redirecting forces from the quadriceps muscle to the
knee joint, which it protects.
Injury. The injury of the lower extremity, Fig. 9.35, is discussed in part in L EVINE
[1993]. The most frequent and severe injuries of the knee-thigh-hip complex are skele-
tal bone fractures and hip dislocations. Fractures of the skeletal bone can be classified
roughly in shaft or diaphysis fracture of the long bones and in crushing or compression
of the short bones and of the meta- and epiphysis (articular surface) near the articula-
tions of the long bones. The injuries can range from dislocations to closed fractures to
severe open fractures.
In front collision, knee or pelvic dislocation can occur if contact is made between
the femur and the dash board, Fig. 9.35(a), (b), (d) (Profs. Choi, Poitout) and (e)
(http://www.sicot.org/: Library: Online Report E006, February 2002, Figure 1 by:
C OSTA -PAZ , R ANALLETTA , M AKINO , AYERZA and M USCOLO [2002]). Because of
crash forces transmitted along the axis of the femur to the hip joint, knee injuries, such
as femoral condyle split and patella fracture, Fig. 9.35(d) and (e), would more likely
occur with relatively harder surfaces of the dash board, while hip dislocation, (b), and
femur shaft fracture, (c), occur with softer knee padding.
Hip dislocation, Fig. 9.35(b) and femoral neck and acetabulum fractures, (f),
are typically seen in a collision of the pelvis with lateral components, causing
the femoral head to punch through the acetabulum through direct force applica-
tion via the greater trochanter (Fig. 9.34(a)). These injuries can occur as a result
of direct contact with side structures. This produces a more complex injury pat-
F IG . 9.35. Lower extremity injury. (Insets (b), (c) and (f) left-hand side: Courtesy Prof. D. Poitout, Ser-
vice de Chirurgie Orthopédique et Traumatologique, Hôpital Nord, Marseille; Inset (d): Private photograph
(Prof. H.Y. Choi); Inset (e): Reproduced by permission of SICOT Sociéte Internationale de Chirurgie Ortho-
pedique et de Traumatologie; Inset (f) right-hand side: Reproduced by permission of Glacier Valley Medical
Education.)
tern, as shown in Fig. 9.35(f) (Prof. D. Poitout and http://www.glaciermedicaled.

com/bone/bonesc11p3.html), that not only involves the articular surfaces of the acetabu-
lum, but also the iliac wing and the pubic rami (arch-like structures) of the pelvis. While
bony disruption of the lower limb constitutes a severe injury, additional complications
can result from the possibility of soft tissue injury.
Model structure and calibration. The modeling aspects of the human lower extremity
are summarized in Fig. 9.36. The H-LE model mainly consists of deformable and dam-
ageable bones, including the pelvis bones, and flesh padding. Active and passive muscle
forces are modeled with bar elements using the Hill muscle model. Nonlinear contact
interfaces model the cartilage layers on hip and knee joints. Fig. 9.36(a) shows the
H-LE model in its context, with the attached muscle bars and with the flesh paddings.
Fig. 9.36(b) shows the modeling of the cartilage layers of the hip joint with nonlinear
contact interfaces. Fig. 9.36(c) contains the modeling of the major hip joint ligaments,
which are modeled with nonlinear bars, while Fig. 9.36(d) represents the major knee
ligaments. The materials are chosen similar to the materials of the upper extremity.
Validations. From knee bolster impact in car frontal crash events, Fig. 9.37 summa-
rizes an investigation made on the injury of the femur (H AYASHI , C HOI , L EVINE ,
YANG and K ING [1996]). Fig. 9.37(a) shows the H-LE model response with distal femur
fracture: (“condyle split”) due to a hard knee bolster padding material. In that case, the
impact force magnitude is high and leads to the observed injury mode. In Fig. 9.37(b)
a soft knee bolster padding material was applied, which leads to the observed hip dis-
location, because the femur axial force magnitude is too low to cause fracture, but too
high over too much time for the hip joint to remain intact. Fig. 9.37(c) represents the
same impact using a knee bolster padding material of intermediate stiffness, that leads
to the observed femur shaft fracture. The latter injury is deemed to be the least damag-
ing and best healing femur/hip injury. Fig. 9.37(d), finally, compares experimental and
calculated impact force time histories.
For a side impact scenario, Fig. 9.38 shows the displacements and the deformations
of the pelvis (upper pictures) and the pelvic bone with von Mises stress contours and
fracture of the pelvic bone (lower pictures).
References on the H-LE model. The H-LE model was built under consultancy
of the following references: BACH , H ULL and PATTERSON [1997] on the measure-
ment of strain in the anterior cruciate ligament; B EDEWI , M IYAMOTO , D IGGES and
B EDEWI [1998] on the human femur FE impact and injury analysis; C AVANAUGH ,
WALILKO , M ALHOTRA , Z HU and K ING [1990] on the biomechanical response and
injury tolerance of the pelvis in side impact tests; C ESARI , B ERMOND , B OUQUET
and R AMET [1994] on testing and simulation of impacts on the human leg; DALSTRA
and H UISKES [1995] on the load transfer across the pelvic bone; D OSTAL [1981] on
a 3D biomechanical model of the hip musculature; E NGIN [1979], E NGIN and C HEN
[1988a], E NGIN and C HEN [1988b] on the kinematics and passive resistances of the
hip joint; F UKUBAYASHI and K UROSAWA [1980] on the contact area and the pressure
distribution of the knee joint; H AYASHI , C HOI , L EVINE , YANG and K ING [1996] on
F IG . 9.36. H-LE model overview.

F IG . 9.37. H-LE model knee bolster impact validations (test from H AYASHI , C HOI , L EVINE , YANG and
K ING [1996]).
the experimental and analytical study of frontal knee impact; K AJZER [1991] on the
impact biomechanics of knee injury; K APANDJI [1974b] (lower limb joint physiology);
K ING [1993] on the injury of the thoraco-lumbar spine and pelvis; K RESS , S NIDER ,
P ORTA , F ULLER , WASSERMAN and T UCKER [1993] on the human femur response
to impact loading; L EVINE [1993] on injury to the extremities; M ARTIN and T HOMP -
SON [1986] on Achilles tendon rupture; M ARTINI , T IMMONS and TALLITSCH [2003]:
Human Anatomy; M OMERSTEEG , B LASKEVOORT, H UISKES , KOOLOOS and K AUER
[1996] on the mechanical behaviour of human knee ligaments; N YQUIST, C HENG ,
E L -B OHY and K ING [1985] on the tibia bending strength and response; PATTIMORE ,
WARD , T HOMAS and B RADFORD [1991] on the nature and causes of lower limb in-
F IG . 9.38. H-LE model side impact validations (pelvis deformation and stress and fracture locations).
juries in car crashes; P ORTIER , T ROSSEILLE , L E C OZ , L AVASTE and C OLTAT [1993]

on lower leg injuries in real-world frontal car accidents; P UTZ and PABST [2000]:
Sobotta Atlas of Human Anatomy; R ENAUDIN , G UILLEMOT, P ÉCHEUX , L ESAGE ,
L AVASTE and S KALLI [1993] on a 3D FE model of the pelvis in side impacts;
ROHEN and YOKOCHI [1983]: Color Atlas of Anatomy; S TATES [1986] on adult oc-
cupant injuries of the lower limb; W YKOWSKI , S INNHUBER and A PPEL [1998] on
a finite element model of the human lower extremity in frontal impact; YANG and
L ÖVSUND [1997] on a human model for pedestrian impact simulation. The following
web sites were consulted: http://www.rad.washington.edu/RadAnat/pelvis.html;
http://www.rad.washington.edu/RadAnat/knee.html; http://www.sicot.org/; http://www.
glaciermedicaled.com/bone/bonesc11p3.html.
F IG . 9.39. Anatomy of the human foot/ankle complex. (Insets (a)–(c): Reproduced by permission of Michael
L. Richardson, University of Washington Medical Center, Department of Radiology; Insets (d) and (e): Re-
produced by permission of Lippincott, Williams & Wilkins.)
9.8. H-Ankle&Foot
Overviews on the H-Ankle&Foot model structure and applications are given in B EAU -
GONIN , H AUG , M UNCK and C ESARI [1995], B EAUGONIN , H AUG and C ESARI
[1996], B EAUGONIN , H AUG , M UNCK and C ESARI [1996], B EAUGONIN , H AUG and
C ESARI [1997].
Anatomy. The anatomy of the human foot/ankle complex is summarized in Fig. 9.39.
The skeleton of the foot consists in the short bones of the tarsus (a) (http://www.rad.
washington.edu/RadAnat/AnkleMortiseLabelled.html and http://www.rad.washington.
edu/RadAnat/AnkleLaterallLabelled.html), namely the talus (or astragalus), which
establishes the articulated connection between the foot and the leg (tibia, fibula),
the calcaneous, the navicular (or scaphoid), the cuboid and the cuneiform bones,
F IG . 9.40. Foot and ankle injury. (Inset (a): Reproduced by permission of ACFAS American College of Foot
and Ankle Surgeons; Insets (b), (d) and (e): Courtesy Prof. D. Poitout, Service de Chirurgie Orthopédique
et Traumatologique, Hôpital Nord, Marseille; Inset (c): Courtesy of Robert A. Christman, D.P.M., Philadel-
phia, PA.)
and of the metatarsal and the phalanx bones (b) (http://www.rad.washington.edu/

RadAnat/FootLateralLabelled.html), (c) (http://www.rad.washington.edu/RadAnat/
FootAPLabelled.html). Besides the major articulations of the talus bone, the Lisfranc
and the Chopart joints provide minor mobility (c). These bones are connected by nu-
merous ligaments and soft tissues, (d) and (e) (N ORDIN and F RANKEL [1989]).
Injury. Injury of the foot/ankle complex, Fig. 9.40 (see indicated web sites), is dis-
cussed in part in L EVINE [1993]. As much as one third of all surviving vehicle crash
victims sustain lower limb injury, where belt use does not alter significantly the risk.
These injuries are not life threatening, but cause extensive health care cost and long pe-
riods of recovery. Ankle and foot injuries are mainly attributed to foot well intrusion in
frontal crashes. They can be classified as skeletal injury, such as ankle fractures (malle-
olar or bimalleolar fractures; tibial pylon fractures; talar fractures) and foot fractures
(e.g., the metatarsal bones; the calcaneus; the cuboid) and internal injury, such as joint
injury (Lisfranc and Chopart joints), ligamentous injury (sprains, tears), tendon injury
(Achilles tendon damage or rupture).
M ORGAN , E PPINGER and H ENNESSEY [1991] described six mechanisms which
they consider the most frequent in vehicle crash (leg trapped between floor and instru-
F IG . 9.41. H-Foot&Ankle model overview (B EAUGONIN , H AUG , M UNCK and C ESARI [1995], B EAUGO -
NIN , H AUG and C ESARI [1996]). (Reproduced by permission of The Stapp Association.)
ment panel with pocketing of instrument panel; contact with foot controls; wheel well
intrusion; contact with floor; collapse of leg compartment; foot trapped under pedals).
These mechanisms correspond to a combination of ankle/foot simple movements like
dorsiflexion, plantarflexion, pronation and supination. These movements can be associ-
ated with direct or indirect loading conditions.
Model structure and calibration. The foot/ankle complex (H-Ankle&Foot) of the

H-Model mainly consists of deformable and damageable bone models with ligaments
modeled with membranes and bars, Fig. 9.41 (B EAUGONIN , H AUG , M UNCK and C E -
SARI [1995], B EAUGONIN , H AUG and C ESARI [1996]).
F IG . 9.42. H-Foot&Ankle model ligament calibration (after PARENTEAU , V IANO and P ETIT [1996] and
B EGEMAN and A EKBOTE [1996]). (Inset (a): Reproduced by permission of ASME; Inset (b): After material
received from Prof. Begeman at Wayne State University, Bioengineering Center.)
F IG . 9.43. H-Foot&Ankle validation: Static crush behaviour: Experimental tests and numerical simulations
(M ASSON , C ESARI , BASILE , B EAUGONIN , T RAMECON , A LLAIN and H AUG [1999]). (Reproduced by per-
mission of IRCOBI.)
F IG . 9.44. H-Foot&Ankle validation: In vitro dynamic impacts: Experimental tests and numerical sim-
ulations (B EAUGONIN , H AUG and C ESARI [1996], B EAUGONIN , H AUG , M UNCK and C ESARI [1996],
B EAUGONIN , H AUG and C ESARI [1997], B EGEMAN and K OPACZ [1991], B EGEMAN , BALAKRISHNAN ,
L EVINE and K ING [1992], B EGEMAN , BALAKRISHNAN , L EVINE and K ING [1993]). (Insets (a)–(g) and (i):
Reproduced by permission of The Stapp Association; Inset (h): ESI Software.)
Nonlinear contact interfaces model the cartilage layers of the major foot/ankle joints.
Nonlinear joint elements connect the lesser bones. Fig. 9.41(a) and (b) show the
H-Ankle&Foot model globally and with its major bones in detail. The latter are mod-
eled with damageable thin shell elements for the cortical bone and with crushable foam
solids for the trabecular bone.
Fig. 9.42 shows a typical ligament calibration test, where a ligament is isolated with
its bony insertions, potted in grips and tested in tension (PARENTEAU , V IANO and
P ETIT [1996]). Typical force-displacement and force-time response curves are shown
(B EGEMAN and KOPACZ [1991], B EGEMAN , BALAKRISHNAN , L EVINE and K ING

[1992], B EGEMAN , BALAKRISHNAN , L EVINE and K ING [1993]), which exhibit un-
loading hysteresis and fiber failure. This behaviour is easily simulated using bars with
nonlinear elastic-plastic material behaviour.
Validations. For static crush behaviour of the H-Foot&Ankle model, Fig. 9.43 shows
experimental tests and numerical simulations, where different calcaneum and talus bone
failure modes are obtained when the material properties of the cortical and trabecular
bones are subject to parametric variations (M ASSON , C ESARI , BASILE , B EAUGONIN ,
T RAMECON , A LLAIN and H AUG [1999]).
For dynamic plantar impacts, Fig. 9.44 depicts the in vitro experimental tests (B EGE -
MAN and K OPACZ [1991], B EGEMAN , BALAKRISHNAN , L EVINE and K ING [1992],
B EGEMAN , BALAKRISHNAN , L EVINE and K ING [1993]) and numerical simulations
for the H-Foot&Ankle model validation (B EAUGONIN , H AUG , M UNCK and C ESARI
[1995], B EAUGONIN , H AUG and C ESARI [1996], B EAUGONIN , H AUG , M UNCK and
C ESARI [1996], B EAUGONIN , H AUG and C ESARI [1997]).
While Fig. 9.44(a)–(h) treat the dorsiflexion case, Fig. 9.44(i) treats the cases of in-
version and eversion due to plantar impacts (B EGEMAN , BALAKRISHNAN , L EVINE
and K ING [1993], B EAUGONIN , H AUG and C ESARI [1996]). The validations are de-
scribed in more detail in the cited literature. The comparisons with the response curves
of a Hybrid III mechanical crash dummy model, indicated in the diagrams, shows a
considerable deviation from the response curves obtained with the human model, which
demonstrates the limited biofidelity of crash dummies.
References on the H-Ankle&Foot model. The following references were consulted for
the establishment of the H-Foot&Ankle model: ATTARIAN , M C C RACKIN , D E V ITO ,
M C E LHANEY and G ARRETT [1985] on the biomechanical characteristics of the hu-
man ankle ligaments; B EGEMAN and KOPACZ [1991], B EGEMAN , BALAKRISHNAN ,
L EVINE and K ING [1992], B EGEMAN , BALAKRISHNAN , L EVINE and K ING [1993]
on the biomechanics of the human ankle impact in dorsiflexion (1991), on the hu-
man ankle dynamic response in dorsiflexion (1992) and for inversion/eversion (1993);
C ESARI , B ERMOND , B OUQUET and R AMET [1994] on testing and simulation of im-
pacts on the human leg; K APANDJI [1974b] (lower limb joint physiology); L ESTINA ,
K UHLMANN , K EATS and M AXWELL A LLEY [1992] on fracture mechanisms of the
foot/ankle complex in car accidents; L UNDBERG , G OLDIE , K ALIN and S ELVIK [1989],
L UNDBERG , S VENSSON , B YLUND , G OLDIE and S ELVIK [1989] on the kinemat-
ics of the ankle/foot complex in dorsiflexion (a) and in pronation and supination (b);
NAHUM , S IEGEL , H IGHT and B ROOKS [1968] on lower extremity injuries in front
seat occupants; N ORDIN and F RANKEL [1989] on the Basic Biomechanics of the
Musculoskeletal System; OTTE , VON R HEINBABEN and Z WIPP [1992] on the biome-
chanics of ankle/foot joint injury; PARENTEAU and V IANO [1996] on the kinematics
and PARENTEAU , V IANO and P ETIT [1996] on the biomechanical properties of the an-
kle subtalar joints in quasi-static loading up to failure; PATTIMORE , WARD , T HOMAS
and B RADFORD [1991] on the nature and causes of lower limb injuries in car crashes;
P ORTIER , T ROSSEILLE , L E C OZ , L AVASTE and C OLTAT [1993] on lower leg in-
juries in real-world frontal car accidents; S TATES [1986] on adult occupant injuries
of the lower limb; W ILSON -M AC D ONALD and W ILLIAMSON [1988] on severy ankle
joint injuries; W YKOWSKI , S INNHUBER and A PPEL [1998] on a finite element model
of the human lower extremity in frontal impact. The following web sites were con-
sulted: http://www.rad.washington.edu/RadAnat; http://www.acfas.org/brankltr.html;
http://astro.ocis.temple.edu/~rchristm/TeachFiles/Fractures/AvFxTib.JPG.
10. The fifth percentile female H-Model
Model structure and calibration. The fifth percentile female H-Model (weight 50 kg;
height 1.52 m) is being built according to the same basic principles than the fiftieth
percentile male H-Model (weight 75.5 kg; height 1.75 m), where the anatomy of the
body was adjusted to the dimensions of the fifth percentile woman, Fig. 10.1. Inset (a)
compares the shape of the model to the shape of the 5th percentile female Hybrid III
mechanical dummy model. Inset (b) gives an overview of the model of the thorax. In-
set (c) gives details of this model, concerning the deformable neck, breasts, rib cage,
thoracic organs and the heart. All sub-models were built and calibrated in a way simi-
lar to the corresponding 50th percentile male sub-models, described earlier. The breasts
were modeled as “bio-bags”, Appendix D, which consist in a non-linear elastic outer
skin with a quasi-incompressible filling in the volume created by the inner thorax lining
and the outer skin. The modeling of the breasts is not considered final at that stage,
and models using solid finite elements for the bulk (fatty tissue) and membranes for the
outer envelope (skin) are built.
Calibration for thorax pendulum impact. Fig. 10.2(a) shows the Kroell pendulum
impact test setup for female cadavers (K ROELL , S CHNEIDER and NAHUM [1971],
K ROELL , S CHNEIDER and NAHUM [1974]). In inset (b) the measured, idealized, high
velocity 6.71 m/s (solid lines) and low velocity 4.27 m/s (dashed lines) impact force
F IG . 10.1. Fifth percentile deformable female H-Model.
versus chest compression corridors for pendulum impacts of 23.15 kg mass are shown.
Inset (c) demonstrates the final calibrated response of the female model at high impact
velocity (6.71 m/s) to fall well within the test corridor. Inset (d) demonstrates at low
impact velocity (4.27 m/s) the influence of various stiffness parameters, and notably of
the presence and absence of the breasts during calibration.
Inset (e) contains sections at 10 millisecond intervals of the compression of the chest
and organs, while inset (f) shows overall pictures and details on predicted rib fracture
during this test.
During the calibration process, a first model used raw material data from the male
model, leading to Fig. 10.2(d), curves (1a) and (1b). This model was generally too
stiff. A preliminary evaluation of the modeling of the breasts with that model showed
a considerable influence of their presence or absence. If breasts were modeled, curve
(1a), the chest deformations were under-predicted, as compared to the case when the
breasts were removed, curve (1b). After calibration of all material data, the response of
the model with breasts was close to the measured test corridors, curve (2).
Validation for out-of-position airbag inflation. In Fig. 10.3(a) the fifth percentile
female H-Model is exposed to a typical “out-of-position” (OOP) driver side airbag infla-
tion scenario, cf. RUDOLF, F ELLHAUER , S CHAUB , M ARCA and B EAUGONIN [2002].
Contrary to a standard driver position (Fig. 10.1 insets (a), (b)), Fig. 10.3 shows the
F IG . 10.2. Calibration of the 5th percentile deformable female H-Thorax (tests after K ROELL , S CHNEIDER
and N AHUM [1971], K ROELL , S CHNEIDER and N AHUM [1974]). (Insets (a), (b): Reproduced by permission
of The Stapp Association.)
F IG . 10.3. Validation of the 5th percentile deformable female H-Model (OOP case) (RUDOLF, F ELLHAUER ,
S CHAUB , M ARCA and B EAUGONIN [2002]).
5th percentile female H-Model positioned close to the steering wheel, being rejected to
the rear by the inflating driver side airbag, inset (c). Whereas in a normal position the
occupant comes into contact with a fully inflated airbag, the OOP occupant is hit by
the airbag while it inflates, which can case injury through the phenomenon called “bag-
slap”. Further, the inflating bag may be trapped under the chin, inset (b), and thus project
the head and neck towards the rear with an over extension of the neck. The mid-sagittal
section plots of inset (b) suggest that injury of the head/neck/thorax complex can be
assessed much more readily from the human model than from the dummy model. Using
human models and models of the standard 5th percentile female Hybrid III mechanical
dummy, one can appreciate the differences between the response of the dummy and the
human, inset (d). Inset (e), finally, gives an impression on the influence of the modeling
if the outer skin of the breasts, which, when not correctly stiffening at large skin strain,
may be reacting in an overly soft manner.
In the out-of-position airbag validation test case, the outer skin of the breasts was
initially modeled with the isotropic material properties as calibrated from the pendulum
impact tests. This led to large deformations in the OOP case, as observed in Fig. 10.3(e)
(left). Such deformations were not seen previously in the pendulum impact simula-
tions, because the pendulum impacted in a more concentrated fashion, between the
breasts. Then the skin was re-modeled with an orthotropic fiber reinforced material,
with stiffening fibers at larger strains, as it is characteristic of skin (YAMADA [1970]
and Fig. 1.1(h)). If this stiffening effect was neglected, the skin was too soft and the
final deformations were too large. It was found that the volume and the mechanical
resistance of the breasts can play an important role in the energy absorption characteris-
tics of the thorax from frontal impacts. It is therefore important to further evaluate their
influence in future studies.
Further investigations on the small female dummy and cadaver tests under static
airbag out-of-position deployment tests can be found, e.g., in C RANDALL , D UMA ,
BASS , P ILKEY, K UPPA , K HAEWPONG and E PPINGER [1999].
A PPENDIX A
Basic Theory of Crash Codes

The following paragraphs are based mainly on Pam–Crash documentation of ESI Soft-
ware. See also H AUG , C LINCKEMAILLIE and A BERLENC [1989a], H AUG , C LINCKE -
MAILLIE and A BERLENC [1989b].
A.1. Overviews on solution methods and finite elements
Solution method overview. Modern crash and impact simulation codes are three-
dimensional, Lagrangian, finite element, explicit and implicit, vectorized and multi-
tasked solid codes for the non-linear dynamic and large deformation analysis of solid
structures in the realm of computational structural mechanics (CSM). They analyze
crash phenomena at discrete points in space and time.
Space discretization. Space is discretized with the most often used finite element
methodology, which is based on the displacement method of structural analysis, where
the discrete nodal displacements and rotations constitute the unknowns of the problem.
A comprehensive account on the theory and application of the finite element method in
engineering is given in the textbook by BATHE [1996]. Some solid codes are coupled to
flow codes, which treat problems in the realm of computational fluid dynamics (CFD), in
order to treat fluid-structure interaction (FSI) dominated problems. Most solid codes are
provided with alternative spatial discretization schemes called particle methods, where a
solid, fluid or gaseous medium is represented by mesh-less discrete (mass or integration)
points, such as in the most common mass-point smooth particle hydrodynamics (SPH)
scheme. This permits to solve certain classes of FSI problems with the solid codes
without the necessity to couple with a separate flow code.
The solid codes allow to model 3D structures of arbitrary geometry using solid el-
ements, membrane elements, plate and shell elements, beam and bar elements, and
discrete spring and joint elements. In typical crashworthiness and impact simulations,
plates and shells are used to model thin-walled metal or plastic components. Beams and
bars are used for stiffening frames, wheel suspensions, shafts, special connections or
secondary components. Solid elements may be used for modeling the bulk of crushable
foams. In typical finite element simulations of the human body, plates and shells are
used for the simulation of cortical bone, beams and bars are used for the modeling of
long bones, tendons and muscles, and solid elements are used for the modeling of the
bulk of soft tissues and for the spongeous bone.
363
364 E. Haug et al. A PPENDIX A
Time discretization. The application of the displacement finite element method

leads to the discretized, coupled and nonlinear equations of motion in each displacement
and rotation degree of freedom. These equations can be integrated in the time domain
by using either implicit or explicit methods. Both methods use time discretization oper-
ators which permit to solve for the unknown displacements, rotations, linear and angular
velocities and accelerations of each degree of freedom, at a given discrete point in time,
from the known states of the structure at previous points in time. An overview on the
explicit and implicit solution methods can be found in H UGHES , P ISTER and TAYLOR
[1979].
Standard implicit methods require linearization of the set of nonlinear equations of
motion and lead to sets of coupled algebraic equations which, for non-linear problems,
must be solved at a considered point in time in an iterative fashion, in order to achieve
dynamic equilibrium at that time. The degree of coupling of the equation set is mea-
sured by the “bandwidth” of the linearized system matrix, which envelopes the extent
of non-zero matrix elements of each of its rows. In a “diagonal” equation matrix the
bandwidth is minimal (one) while in a “full” matrix the bandwidth is maximal (number
of equations). Implicit methods can be used together with explicit solution, when the
explicit solution is costly, such as in quasi-static problems, which arise, for example,
when seating an occupant model into a deformable car seat before simulating a crash
scenario, or by calculating the elastic spring-back of stamped parts.
Standard explicit methods do not require repeated solution of linearized coupled
equation systems, and lead to a set of uncoupled algebraic equations when a lumped
(diagonal) mass matrix is assumed. Solution of diagonal equation systems is trivial and
computer time per discrete solution time step is negligible as compared to the computer
time needed to repeatedly solve the potentially huge coupled systems of algebraic equa-
tions of the implicit methods. To date, crash models with up to one million thin shell
finite elements and more have been built and solved successfully on powerful computers
with the explicit methods.
The time increment of explicit methods, however, is restricted for solution stability,
while, in principle, the time increment in unconditionally stable implicit methods is not
restricted in size. In typical crashworthiness and impact studies over relatively short
durations and involving large distortions of the structural parts, this advantage of the
implicit methods has no bearing, however, because the structural states must be known
at many discrete points in time in order to allow for an accurate tracing of the complex
physical phenomena, including “contact”, and to account for material and geometrical
non-linearity. Time increments of the order of one microsecond and less are typical in
crash simulations with explicit time integration.
In crashworthiness and in higher velocity dynamic impact studies, therefore, the ex-
plicit time integration methods have proven computationally advantageous. Implicit
solution algorithms, on the other hand, may be applied with advantage to quasi-static or
slow vibration dynamic problems with limited non-linearity.
Finite elements overview. Standard finite elements are defined by a set of nodes and
a connectivity array, which relates their geometric topology to these nodes. In nonlin-
ear crash and impact analysis, experience shows that “higher order” elements are not
S ECTION A.1 Basic theory of crash codes 365
beneficial. This is mainly due to the fact that sharp stationary or traveling plastic hinge
folding lines of crushed thin shell structures, made of elastic-plastic material, are not
well accommodated even by complex interpolation shape functions of higher order ele-
ments, and that plastic hinge lines therefore require a dense nodal point spacing in the
finite element models of thin walled structures, even when computationally expensive
higher order elements are used.
Similarly, high gradients of three-dimensional stress and of plastic deformation are
better accommodated by dense finite element models of simple 8-node 3D solid ele-
ments, than by using higher order solids that, for computational competitiveness, must
have coarser meshes. Similar remarks apply for the simulation of fracture, where the
small scale of the physics of fracture requires even higher mesh densities, sometimes
achieved by automatic adaptive local mesh refinement schemes. Recently, so-called
“mesh-less” methods are applied to simulate fracture, such as the EFG (element-free
Galerkin), DE (discrete elements), SPH (smoothed particle hydrodynamics) (M ON -
AGHAN and G INGOLD [1983] and M ONAGHAN [1988]), FPM (finite point methods),
etc. These advanced space discretization methods are presently under development and
they are implemented in commercial crash codes. One instance of recent trial appli-
cations in biomechanics of the mesh-less SPH methods is the simulation of coupled
structure fluid interaction (FSI) problems, where the fluid is simulated with particles,
enclosed in a deformable organ modeled with finite elements. Another emerging method
is the coupling of solid codes and flow codes for the simulation of FSI events (L ÖHNER
[1990]), such as the rupture of the blood-filled aorta in a chest impact scenario.
In many structures, the most important crash simulation finite element is the thin shell
element. The most used thin shell element is a bilinear four node quadrilateral element,
based on the Mindlin–Reissner plate theory. One of the most efficient thin shell elements
was originally developed by B ELYTSCHKO and T SAY [1983] and B ELYTSCHKO and
L IN [1984]. The Mindlin–Reissner plate theory takes the transverse shear deformation
of the plate into account and it presumes that lines normal to the plate mid-surface
remain straight, but not necessarily normal.
In the classical Kirchhoff–Love plate theory the normal lines to the plate mid-surface
remain both, straight and normal, and the influence of the transverse shear deformation
is neglected. The implementation of this theory requires slope-compatibility across el-
ement edges (C 1 -continuity), resulting in a complex finite element formulation. On the
other hand, Mindlin–Reissner theory requires only C 0 -continuity in the shape functions
for assuring complete inter-element deformation compatibility. This greatly simplifies
the FE-formulation.
In the case of the Belytschko element, a reduced domain integration technique with
one-point quadrature is applied for calculating the nodal forces contributed by the
plate and shell elements. This technique avoids membrane-locking, but permits certain
zero-energy or kinematic deformation modes of the elements, with which no resisting
material stresses are associated. Such zero-energy modes are called hourglass-modes,
which, if excited, can lead to numerical instability through uncontrolled spurious os-
cillations. In order to prevent hourglass modes, a built-in hourglass control algorithm
is implemented, as developed by B ELYTSCHKO , W ONG , L IU and K ENNEDY [1984].
It effectively avoids the numerical instability problem associated with one-point quadra-
ture shell elements, by effectively damping out the zero-energy modes.
A.2. Explicit solution method outline
Three dimensional structures. In a three-dimensional solid model there are 3 transla-

tional degrees-of-freedom (DOF) per node. In a three-dimensional thin shell numerical
model there are 6 degrees of freedom per nodal point, i.e., 3 translations and 3 rota-
tions. In a three-dimensional thin plate numerical model there are 5 local degrees of
freedom, i.e., 3 translations and 2 rotations. Crash models are mainly spatial models of
elements that can be subdivided into layers with plane stress conditions (plate and thin
shell theory).
Lagrangian discretization. This term refers to a choice of independent variables for

the problem. In the Lagrangian formulation, each material particle is characterized by its
initial coordinates (x0 , y0 , z0 ) and its actual coordinates are chosen to be the dependent
variables of the problem,
x = x(x0 , y0 , z0 , t),
y = y(x0 , y0 , z0 , t),
z = z(x0 , y0 , z0 , t).
In this formulation, the ordinary differential equations, obtained after spatial discretiza-
tion, describe the dynamic equilibrium of the material particles which stand for the
original continuum. For example,
M d2 x/dt 2 + C dx/dt + Kx = Fext ,
are the discretized equations of motion in the linear case, where M is the mass matrix,
C is the damping matrix, K is the stiffness matrix and Fext is the vector of applied loads.
The acceleration of a particle is equal to the material derivative of the velocity, v, and
equal to the partial derivative of the velocity because x0 , y0 , z0 are constants:
dv ∂v ∂v ∂x0 ∂v ∂y0 ∂v ∂z0 ∂v
= + + + =
dt ∂t ∂x0 ∂t ∂y0 ∂t ∂z0 ∂t ∂t
The discrete finite element mesh points coincide with material points and have time
dependent coordinates. The finite element mesh will thus deform with the material. This
will engender distortions of the mesh, which can result in decreasing stable solution time
steps and in solution inaccuracies due to excessive element distortions.
Finite elements. Finite element “unstructured” meshes permit to establish numerical

models of a physical structure with realistic discretization and representation of the
boundary conditions
Only the simplest finite elements are used in the PAM-Crash program because it is
believed that fine meshes of simple elements give better results in highly distorting
structures than coarse meshes of high-order elements. The most often used finite ele-
ments in crash codes are
– 4–8 node solids (1 point and fully integrated);

– 3–4 node bilinear shells (1 point and fully integrated);
– 3–4 node isoparametric membrane elements (fully integrated);
– 2 node beam, bar and joint elements.
The optimization of industrial crash codes has focused on the treatment of the most
frequently used shell element.
Explicit integration scheme. A vibrating spring/dashpot-mass system consists of a

mass, m, a dashpot with constant c, a spring with constant k, and an external load,
f (t), Fig. A.1.
Consider the second order ordinary differential equation which expresses dynamic
equilibrium of this system, where x is the displacement, ẋ is the velocity c is the damp-
ing coefficient, ẍ is the acceleration and m is the mass,
mẍ + cẋ + kx = f (t) (1 DOF).

The central difference scheme considers the following time-axes around a discrete
point in time, tn , Fig. A.2.
The known quantities are the displacement xn at time tn , and the velocity ẋn−1/2 at
the intermediate time tn−1/2 . The wanted quantities are the displacement xn+1 at time
tn+1 , and the velocity ẋn+1/2 at time tn+1/2 . Dynamic equilibrium at time tn is expressed
F IG . A.1. Spring/dashpot-mass system.
F IG . A.2. Central difference scheme.

as
mẍn = fn − cẋn − kxn .

Since all terms on the right hand side are known, one can first solve for the accelera-
tion at time tn , ẍn , and then apply the central difference time integration, Fig. A.2, and
solve for the unknown quantities, ẋn+1/2 at time tn+1/2 and xn+1 at time tn+1 as follows

ẍn = m−1 fn − cẋn − kxn ,

ẋn+1/2 = ẋn−1/2 + tn ẍn ,
xn+1 = xn + tn+1/2 ẋn+1/2 .
Stable time step. The criterion for solution stability of the central difference explicit
time integration can be derived formally as follows (BATHE [1996]). For a 1-DOF free-
vibration system the equation of motion becomes
ẍn + 2ξ ωẋn + ω2 xn = 0 (free vibration),

√ √
where ξ = c/ccrit is the damping ratio, ccrit = 2km, is the critical damping, ω = k/m
is the circular frequency of vibration, and k and m are the spring stiffness and vibrating
mass, respectively. The central difference scheme dictates, using constant time step, t,
around time tn ,
ẋn+1/2 = (xn+1 − xn )/t,

ẋn−1/2 = (xn − xn−1 )/t,
ẍn = (ẋn+1/2 − ẋn−1/2 )/t = (xn+1 − 2xn + xn−1 )/t 2 .
Substitution of these expressions into the above equation of motion yields the following
recursive system of equations
2−ω2 t 2 ωt
xn+1 − 1−ξ xn
= 1+ξ ωt 1+ξ ωt .
xn 1 0 xn−1
For zero damping (ξ = 0) this becomes

xn+1 2 − ω2 t 2 −1 xn
= , i.e.,
xn 1 0 xn−1
xn+1 = Axn .
In general, one can write the equivalent recursive matrix relationship
xn+1 = Axn = A2 xn−1 = · · · = An x1 ,

where the operator matrix A can be seen to recur in powers of n. This relationship is
stable only if An remains bounded for all values of n. This means that the absolute
values of the eigenvalues of A, or its “spectral radius”, ρ, must be less than or equal to
one, i.e.,
ρ(A) = max |λi | 1 for stability.

F IG . A.3. Free vibration bar element.
In case of zero damping, the eigenvalues of A follow from Ax = λx, with √ solutions
det(A − λI) = 0, where I is the 2 × 2 unit matrix. This leads to λ1,2 = a ± a 2 − 1,
in which a = (2 − ω2 t 2 )/2. Enforcing |λ1,2 | = 1 one has λ21,2 = 1, from which fol-
lows that a 2 = 1 satisfies the condition |λ1,2 | = 1 non-trivially. From this equation, the
condition for stability of the time step is given by
stable t 2/ω.
For the case of the freely vibrating spring/mass system, stability is enforced if

t 2 m/k.
E XAMPLE . Consider the free vibration 1 DOF system, made of a bar element with
uniform mass and stiffness distribution, with half length, L/2, elastic modulus, E, mass
density per unit volume, ρ, and cross section area, A, as shown in Fig. A.3.
The criterion for stability becomes in this case

m ρ L L
tn < 2 =L =√ = ,
k E E/ρ c
√
where c = E/ρ is the speed of sound in the bar material. For steel one has roughly
c = 5 km/s = 5 mm/microsecond. This means that for a shell element with dimensions
of 5 × 5 mm the stable time step will be about 1 microsecond, which corresponds to the
time it takes of an acoustic signal to travel across the element.
Implicit integration outline. Consider the second order differential equation (c = 0)
mẍ + kx = f
and a discretized time axis as follows,
Let the known quantities at time tn be xn and ẋn . The wanted quantities are then ẋn+1
and xn+1 . Dynamic equilibrium at time tn+1 is expressed as
mẍn+1 + kxn+1 = fn+1 ,

where xn+1 is unknown. Direct solution for ẍn+1 or xn+1 is therefore impossible. One
can now apply forward differences and substitute
ẋn+1 = (xn+1 − xn )/t,

ẍn+1 = ẋn+1 − ẋn /t

and thus:
ẍn+1 = (xn+1 − 2xn + xn−1 )/(t)2 ,

which after substitution yields

m/t 2 + k xn+1 = fn+1 − m/t 2 (2xn − xn−1 ).

This equation can be solved for the unknown displacement at time tn+1 ,

−1
xn+1 = m/t 2 + k fn+1 − m/t 2 (2xn − xn−1 ) .

The other variables are then obtained as
ẋn+1 = (xn+1 − xn )/t,

ẍn+1 = ẋn+1 − ẋn /t.

This scheme works independently of the chosen value of t. It is said to be “uncon-
ditionally stable”. It necessitates, however, to perform the solution of (m/t 2 + k)−1 ,
which is in general a costly operation because the stiffness matrix is not diagonal. For
large time steps, however, the solution may remain stable, but one may accumulate pe-
riod elongation and amplitude decay time integration errors in freely vibrating systems.
For non-linear equations of motion, the linearized equations of dynamic equilibrium
must be solved iteratively to ensure dynamic equilibrium at times tn . If this is neglected,
the solution, may become inaccurate and unstable.
Vectorized and multi-tasked codes. Explicit codes are well suited for vectorized
super-computers because elements carry no dependencies, i.e., elements can be treated
simultaneously, no K-matrix is stored and no I/O operations are required. This means
that the elapsed time is practically equal to the CPU time in a computer run. For the
same reasons, explicit codes parallelize well. Most industrial crash codes come with
shared and distributed memory parallel architecture.
A.3. Contact treatment outline
Overview on contact algorithms. The successful and efficient treatment of contact

events is of prime importance in crash and impact simulation. To identify potential
contact surfaces of a given structure, the crash codes allow to define contact interface
entities, which are most often given by collections of surface facets, identical to finite el-
ement surfaces. Actual contact can be defined a distance tcontact (contact thickness) away
from the mid-surface defined by the facets represented by shell elements. Most contact
algorithms are based on a node-to-segment treatment. After preliminary geometrical
TABLE A.1
Some typical contact options (overview)
Group Type∗ Action
Rigid walls one-sided moving or fixed boundary condition
Nodal constraints pairs of nodes are constrained to displace together
Special sliding and tied interfaces Type 1 sliding without separation

Type 2 3 DOF or 6 DOF tied with single failure
Internal solid element contact Type 10 internal solid anti-collapse contact
Safety specific contacts Type 7 multiply self-impacting contact for airbag

Type 11 body-to-plane contact (force-deflection)
Type 12 body-to-body contact (force-deflection)
Type 21 body-to-multiplane contact
(force-deflection or stress-strain)
Type 37 enhanced self-impacting contact for airbag
Recommended contacts for crash Type 31 penalty-free node-to-segment contact

simulation Type 32 penalty or kinematic tied contact with
distance and failure
Type 33 segment-to-segment contact with edge treatment
(3D bucket search)
Type 34 node-to-segment contact with edge treatment
(3D bucket search)
Type 36 self-impacting contact with edge treatment
(3D bucket search)
Type 42 mesh independent spotweld parts
Type 44 node-to-segment contact with smooth contact surface
Type 46 edge-to-edge self-impacting contact
∗ PAM-Crash code.
proximity searches for potential contact events, the so identified candidate contact seg-
ments are fine-checked for mutual penetrations of their nodes and surfaces. Detected
penetrations are limited (“penalized”) or prevented by the algorithms.
In impact biomechanics several kinds of contact events are of importance. On the one
hand, the biomechanical models may contact or collide with external objects (chest-
to-seat belt, body-to-airbag, body-to-car interior, etc.). On the other hand, body parts
and organs may be in mutual contact (arm-to-chest, chin-to-chest, leg-to-leg, internal
organ-to-internal organ, organ-to-wall of body cavity, brain-to-skull, etc.).
Contacts may be sticking (“tied” options), sliding (“slide” options), multiple (“slide-
and-void” options), failing (rupture options), and sliding can be with and without
friction. Each contact event may require a particular treatment, with or without penalty
algorithms. To demonstrate the importance editors of crash codes attribute to con-
tact treatments, Table A.1 lists some often used contact options. The listed rigid wall
and nodal constraint options are simpler contact treatments that can serve for de-
tecting collisions with rigid obstacles (rigid walls) or that can be used to constrain
specified pairs of nodes to move together (nodal constraints). The interested reader is
F IG . A.4. Rigid wall contact.
advised to consult the notes manuals of the commercial crash codes for more detail
(see also: H UGHES , TAYLOR , S ACKMAN , C URNIER and K ANOKNUKULCHAI [1976],
H ALLQUIST, G OUDREAU and B ENSON [1985] and H AUG , C LINCKEMAILLIE and
A BERLENC [1989b]).
In simulations of impact biomechanics all listed options can be applied. Some details
are provided next for the most popular rigid wall and penalty contact options. Contact
types 11 and 12 are often used in the simple rigid multi-body models (HARB models)
and they provide a “soft” penalization of penetrating rigid volumes according to user-
defined penetration restoring force-deflection curves.
Rigid walls. Fig. A.4 shows the principle of rigid wall treatment, where a car may
collide with an inclined rigid wall or where a car occupant may collide with a simplified
model of the car interior, modeled with several rigid wall segments.
The rigid walls are impenetrable and the contacting bodies may slide along the walls
with and without specified friction coefficients. The rigid walls may not be able to move,
or they can be assigned some specified motion. Penetration of nodes into a rigid wall is
prevented by an algorithm that sets the relative normal velocity to zero.
Nodal constraints. Nodal constraints are often used to tie together two non-matching
finite element meshes, as it can arise when two internal organs are modelled indepen-
dently and when the two organs have a common mesh interface. Instead of re-meshing
the interfaces, one may simply identify pairs of nodes to move together.
Penalty methods. If penetration of a node into a segment is detected, it will be pe-

nalized in most contact algorithms by elastic restoring forces of fictitious (non-)linear
springs, which are compressed by the amount of penetration. In dynamic contact events,
dashpots can be added to dissipate energy and to prevent from spurious oscillation.
Fig. A.5 shows two potentially contacting surface segments. The algorithm checks
if a slave node of the right segment touches the left master segment. Then the master
and slave roles are reversed and the algorithm checks if a slave node of the left seg-
ment touches the right master segment. Inverting the master-slave roles prevents from
contacts to be missed.
F IG . A.5. Contact interfaces.
F IG . A.6. Penetration and perforation.
Fig. A.6 shows the application of a penalizing contact spring after “penetration” of
a slave node into a master segment. If the slave node penetrates a distance beyond
the contact thickness, tcontact , over the mid-surface of the master segment, then this
node will have “perforated” the master segment and it will escape to the other side.
Perforations can happen when the impact is too violent to be contained by the penalty
springs.
A PPENDIX B
Biomaterial Mechanical Properties

The following text and tables are extracted mainly from a report of the HUMOS-1
Project (YANG [1998]):
Yang, J. (1998) Bibliographic Study, Report 3CHA/980529/T1/DB, Chalmers Uni-
versity of Technology, SE-41296 Göteborg, Sweden.
The HUMOS project was funded by the European Commission under the Industrial
and Materials Technologies program (Brite–EuRam III). This report has been prepared
with the aim to provide input data for human impact models. It contains many references
to the abundant literature on biomaterials, and the most important data are summarized
hereunder. For simulations in impact biomechanics, the mechanical properties of the
musculo–skeletal system are the most needed and known. The static and dynamic me-
chanical properties of the soft tissues and organs are the least known properties.
The mechanical properties of all living tissues undergo wide scatter through different
ages, gender, biological, physical and loading conditions. Most of the listed properties
delimit the observed scatter and can serve as guide lines and upper and lower bounds in
human model calibration, and in stochastic analyses. Most biomaterials are non-linear,
anisotropic, different in tension and compression and they are deformation rate depen-
dent. The majority of the indicated values represent their average linear behavior. The
properties of bone and soft tissue materials can depend much on the location in the
human body, i.e., vary from member to member and within an individual part. Most
properties were obtained from cadaver tests, which can only approximate their behavior
in living bodies.
While the mechanical properties of the long bones (femur, tibia, fibula, humerus,
ulna, radius) is a well studied subject the in biomechanics, data are much scarcer for
pelvic bone, ribs, sternum, clavicle, scapula, short bones of the feet and other parts. For
this reason the following data are concentrating on the properties of long bones.
B.1. Bone mechanical properties
Quasi-static properties of bone. Table B.1 contains the average quasi-isotropic proper-
ties of the most important skeletal bones, where N is the number of tested specimens,
E is the linear isotropic Young’s modulus, ν is Poisson’s ratio, G is the shear modulus
and σu is the ultimate stress.
Table B.2 shows the orthotropic stiffness matrix coefficients in directions 1, 2 and 3
of femoral cortical bone from 60 specimens (A SHMAN ET AL . [1984]), where axis 1 is
375
376 E. Haug et al. A PPENDIX B
TABLE B.1
Elastic moduli and ultimate stress for bones (overview)
Authors Bone Test (N ) E [Gpa] ν G [Gpa] σu [Mpa]
M C E LHANEY [1966] general quasi-static 20 ± 5 3.15 ± 0.25 16 ± 2.5

cortical strain rate 0.001 15.1 15.0
strain rate 300 29.2 28.0
general compression 0.001–9.8 0.02–2.5
trabecular
vertebral compression 0.01–1.5
bodies
YAMADA [1970] femur tensile 14–18 12–15

S EDLIN ET AL . [1965] cortical compressive 15–19 14–21
S EDLIN ET AL . [1966] shear 15.5 7–8
R EILLY, bending 3.28 16–18
B URSTEIN and torsion 6.8
F RANKEL [1974]
R EILLY ET AL . [1975]
tibia tensile 18–29 12–17
cortical compressive 25–35 18–21
shear 7–8
YAMADA [1970] femur tensile 17.6 12.4

(tensile dry) (20.2) (15.1)
tibia tensile 18.4 14.3
fibula tensile 18.9 14.9
humerus tensile 17.1 12.5
radius tensile 18.5 15.2
ulna tensile 18.8 15.1
average tensile 18.3 14.0
Q UHAN [1989] pelvis bending 63 yr 5.26 ± 2.09

cortical (36)
bending 23 yr 3.76±1.78
(12)
pelvis bending 63 yr 4.16±2.02
trabecular (29)
bending 23 yr 3.03 ± 1.63
(13)
M C E LHANEY [1970] skull radial comp. 2.4 ± 1.4 0.19 ± 0.08 740 ± 350
sandwich tangential comp. 5.6 ± 3.0 0.22 ± 0.11 970 ± 360
tangential tension 5.4±2.9 430±190
G RANIK and S TEIN [1973] ribs bending 11.5±2.1 11±3
S ACRESTE , ribs bending 6.14±4.26 8.6±5.5

B RUN -C ASSAN , FAYON ,
TARRIERE , G OT and
PATEL [1982]
S ECTION B.1 Biomaterial mechanical properties 377
TABLE B.1
(Continued)
Authors Bone Test (N ) E [Gpa] ν G [Gpa] σu [Mpa]
B URGHELE and calcaneous static 2100 [N]

S CHULLER (10 mm/min) (mean force)
[1968] intact bone
dynamic 2620 [N]
(500 mm/min) (mean force)
intact bone
cortical (largest 2890–3200 2.87–3.79 (stress)
over bone); cut
specimens:
2 × 0.5 × 0.5 cm
talus static (10 mm/min) 446 [N]
intact bone (mean force)
dynamic (500 mm/min) 468 [N]
intact bone (mean force)
TABLE B.2
Orthotropic stiffness matrix [Gpa] for human femoral cortical bone
E11 E22 E33 G12 G13 G23 ν12 ν13 ν23 ν21 ν31 ν32
12.0 13.4 20.0 4.53 5.61 6.23 0.376 0.222 0.235 0.422 0.371 0.350
TABLE B.3
Transverse isotropic stiffness matrix for human femoral cortical bone
E11 E33 ν31 ν12 G13
Mean values [GPa] 11.5 17.0 0.46 0.58 3.28

Number of tests (N ) 31 170 147 26 166
Standard deviation [%] 15–20% in tension 30% 30% 10%
7–10% in compression
the radial direction, axis 2 is the circumferential (or transverse) direction, axis 3 is the
longitudinal direction of the cylindrical bone shaft, Eij are the elastic modules, Gij are
the shear modules and νij are Poisson’s ratios.
Table B.3 contains transverse isotropic stiffness matrix coefficients for human
femoral cortical bone (E11 = E22 ) from a population over the age spans of 19–80 years
(R EILLY, B URSTEIN and F RANKEL [1974]), where N is the number of specimens, E11
is the elastic modulus for transverse or radial specimens, E33 is the elastic modulus for
longitudinal specimens, ν12 is Poisson’s ratio for transverse or radial specimens, ν31 is
Poisson’s ratio for longitudinal specimens and G13 is the shear modulus.
Table B.4 contains trabecular orthotropic stiffness properties found in the human
proximal tibia from 3 males (ages 52, 55 and 67) (A SHMAN ET AL . [1986]).
TABLE B.4
Trabecular bone properties from human proximal tibias
Module Average Standard deviation Range
E11 [MPa] 346.8 218 110–1230

E22 [MPa] 457.2 282 140–1750
E33 [MPa] 1107.1 634 340–3350
G12 [MPa] 98.3 66 30–380
G13 [MPa] 132.6 78 35–410
G23 [MPa] 165.3 94 45–460
Density [kg/m3 ] 263.4 135 130–750
TABLE B.5
Trabecular bone properties from human pelvis
E11 E22 E33 G12 G13 G23 ν12 ν13 ν23 ν21 ν31 ν32
Average [MPa] 59.8 57.3 43.2 26.0 22.6 22.6 0.18 0.24 0.21 0.17 0.16 0.14
Standard Dev. 44.9 44.6 39.9 19.1 17.1 17.2 0.11 0.14 0.16 0.10 0.07 0.09
Table B.5 lists the trabecular orthotropic stiffness properties of the pelvis bone (DAL -
STRA ET AL . [1993]).
Tables B.6 and B.7 list further trabecular bone mechanical properties, showing the
great variety of these values.
Strain rate dependent properties of bone. Bone is strain rate dependent material. It
exhibits stiffer and stronger behavior at high strain rate. Typical values found for cortical
are listed in Table B.8.
Age dependent properties of bone. Mechanical properties of human bones change

with age as a consequence of changes of density and mineral content. According to
B URSTEIN ET AL . [1976], the elastic modulus and tensile strength of human bone de-
crease slowly after the age of about 45 years. C URREY [1975] showed that the elastic
modulus and bending strength both increase with age until the age of about 30 years,
but decrease thereafter. The decrease is associated with, and mainly caused by, the in-
creased porosity and demineralization of bone. M C C ALDEN ET AL . [1993] found that
the change in porosity played a greater role in the reduction in strength than the change
in mineral content. Both factors reduce the ability of bones to undergo plastic deforma-
tion before fracture starts. Table B.9 shows important trends of the age dependency of
the mechanical properties of human bone with age. Table B.10 gives an example for the
tensile and compression age dependent mechanical properties of the cortical bone of the
femur. Similar trends with age are found in the other bones.
TABLE B.6
Compressive properties of trabecular bone [MPa]
(a) Proximal femur Test piece E σu
B ROWN [1980] 9.5 mm length 345 –

5 mm cubes – 1.2–3.1
C IARELLI ET AL . [1986], C IARELLI ET AL . [1991] 10 mm length 58–2248 2.1–16.2
8 mm cubes 49–572 –
E VANS ET AL . [1961] 25 × 7.9 mm prisms – 0.21–14.82
M ARTENS ET AL . [1983] 8 mm diameter 1000–9800 0.45–15.6
S CHOENFELD ET AL . [1974] 7.9 mm cubes 20.68–965 –
4.8 mm diameter – 0.15–13.5
(b) Distal femur Test piece E σu
B EHRENS [1974] 5 mm slab – 2.25–66.2

C IARELLI ET AL . [1986], C IARELLI ET AL . [1991] 8 mm length 58.8–2942 –
8 mm cubes 7.6–800 19
D UCHEYNE [1977] 5 mm diameter – 0.98–22.5
(c) Proximal Tibia Test piece E σu
B EHRENS [1974] 5 mm slab – 1.8–63.6

10.3 mm diameter – –
C ARTER and H AYES [1977] 5 mm length 1.4–79 1.5–45
C IARELLI ET AL . [1986], C IARELLI ET AL . [1991] 8 mm cubes – 0.52–11
G OLDSTEIN [1983] 7 mm diameter 8–457 1–13
10 mm length
H VID ET AL . [1985] 10 mm length 4–430 13.8–116.4
5 mm slabs
L INDAHL [1976] 14 × 9 mm
Male 34.6 3.9
Female 23.1 2.2
L INDE [1989] No constraint 113–853
W ILLIAMS ET AL . [1982] 5–6 mm cubes 10–500 1.5–6.7
(d) Vertebral bodies Test piece E σu
A SHMAN ET AL . [1986] 5 mm diameter 158–378 –

10–15 mm length
BARTLEY [1966] Lumbar region – 2.9
G ALANTE ET AL . [1970] 7, 10 mm diameter – 0.39–5.98
L INDAHL [1976] 10 × 9 × 14 mm
Male 55.6 4.6
Female 35.1 2.7
M C E LHANEY [1970] 10 mm length 151.7 4.13
ROCKOFF ET AL . [1969] Lumbar region – 0.69–6.9
S TRUHL ET AL . [1987] 8&6 mm cubes 10–428 0.06–15
W EAVER ET AL . [1966] 10 mm cube – 0.34–7.72
YAMADA [1970] 40–49 years 88.2 1.86
60–69 years 68.6 1.37
(e) Pelvic trabecular Test piece E σu
D ALSTRA ET AL . [1993] 6.5 mm cubes 58.9 –

TABLE B.7
Shear properties of trabecular bone [MPa]
Test piece G σu
A SHMAN ET AL . [1986] 5 mm diameter 58–89

10–15 mm length
TABLE B.8
Cortical bone strain rate dependent properties
Cortical bone (M C E LHANEY [1966]) Compression Compression

(human embalmed) (fresh cow bone)
Strain rate from 0.001 to 300 from 0.001 to 300

E [GPa] from 15.1 to 29.2 from 18.6 to 33
Strength [MPa] from 150 to 280 from 175 to 280
TABLE B.9
Age dependency of bone properties
(a) Cortical bone ultimate tensile strength for the decade

age groups (after Z HOU , Q. ET AL. [1996]).
(Reproduced by permission of The Stapp Association.)
Observations:
(1) The absolute strength values obtained by different au-
thors are different, but the trend with age is similar.
(2) The maximum strength is reached around the 30–39
age group.
(3) The ultimate tensile strength of the elderly groups (70–
89) drops by 15% to 25% compared to the maximum.
(b) Femoral cortical bone ultimate tensile stress for the

decade age groups (after M C C ALDEN ET AL . [1993]).
(Reproduced by permission of The Journal of Bone and
Joint Surgery, Inc.)
The straight line represents the best fit with the use of lin-
ear regression analysis (from 253 specimens excised from
47 cadavers).
(c) Vertebral trabecular bone ultimate stress relation-

ship between the age groups (after M OSEKILDE ET AL .
[1986]).
(Reproduced by permission from Elsevier.)
TABLE B.10
Age dependent femur cortical mechanical properties
(a) Tensile properties of femur Elastic modulus Yield stress Ultimate stress
[MPa] [MPa] [MPa]
B URSTEIN ET AL . [1976] 20–29 years 17000 120 140

30–39 years 17600 120 136
40–49 years 17700 121 139
50–59 years 16600 111 131
60–69 years 17100 112 129
70–79 years 16300 111 129
80–89 years 15600 104 120
(b) Compressive properties of femur Elastic modulus Yield stress Ultimate stress
[MPa] [MPa] [MPa]
B URSTEIN ET AL . [1976] 20–29 years 18100 209

30–39 years 18600 209
40–49 years 18700 200
50–59 years 18200 192
60–69 years 15900 179
70–79 years 18000 190
80–89 years 15400 180
(c) Bending properties of femur Elastic modulus Yield stress Ultimate stress
[MPa] [MPa] [MPa]
YAMADA [1970] 20–29 years 151

30–39 years 174
40–49 years 174
50–59 years 162
60–69 years 154
70–79 years 139
80–89 years 139
(d) Torsion properties of femur Elastic modulus Yield stress Ultimate stress
[MPa] [MPa] [MPa]
YAMADA [1970] 20–29 years 3430 57

30–39 years 3430 57
40–49 years 3140 52.7
50–59 years 3140 52.7
60–69 years 2940 48.6
70–79 years 2940 48.6
80–89 years – 48.6
B.2. Ligament mechanical properties
Ligaments are fibrous materials with pronounced non-linear force-displacement re-

sponse. The mechanical properties of some ligaments are summarized in Table B.11,
where vd is the deflection rate, D is the deflection at rupture, ε is the strain at rupture,
E is the elastic modulus. For reasons of practical feasibility, ligaments are most often
382
TABLE B.11
Some ligament mechanical properties
Specimen type and vd D F ε E [kN/mm2 ] Other properties
conditions (source) [mm/ms] [mm] [kN] [%] (= GPa)
(a) Collagen (V IIDIK [1987]) σu = 0.045–0.120 GPa
(b) Typical force-deformation (I) = toe region;

curve for ligament for monotonic (II) = linear region;
forcing (III) = region of micro-failure;
(F RANK and S HRIVE [1994]) (IV) = failure region.
(Reproduced by permission of At the top are schematic
the University of Calgary.) representations of fibers going
E. Haug et al.
from crimped (I) through
recruitment (II) to progressive
failure (III and IV)
(c) Human cruciate ligament

(N OYES and
G ROOD [1978])
Non-linear stress-strain behavior
(Reproduced by permission of
the Journal of Bone and Joint
Surgery, Inc.)
A PPENDIX B
TABLE B.11
S ECTION B.2
(Continued)

(d) Human maximum

calcaneo-fibular and 0.4 mm/ms
tibio-talar ligaments with cycle load
(B EGEMAN and
A EKBOTE [1996]) Static
and dynamic isolated
bone-ligament-bone test
Biomaterial mechanical properties

pieces
(Reproduced by
permission of Prof.
Begeman, Wayne State
University,
Bioengineering Center.)
(e) Human ankle/foot ligaments (ATTARIAN , M C C RACKIN , D E V ITO , M C E LHANEY and G ARRETT [1985])
(i) Anterior tibio-fibular 1.50E−04 3 0.016 (m)
ligament 1.50E−02 0.02 (m)
Cyclic load to isolated 1.00E+00 0.042 (m)
bone-ligament-bone
test piece
Effect of strain rate
(Reproduced by
permission of
Foot & Ankle Intl.)
383
384
TABLE B.11
(Continued)

(ii) Calcaneo- 1.50E−04 3 0.029 (m)

fibular ligament 1.50E−02 0.031 (m)
Effect of strain rate 1.00E+00 0.087 (m)
(Reproduced by
permission of
Foot & Ankle Intl.)
E. Haug et al.
(iii) Anterior 1.01 ± 0.07 5.1 ± 0.5 (f) 0.1389 53 3.999E−02 maximum load = force required to completely
tibio-fibular ±0.0235 (max) ±6 (f) ±0.854E−02 disrupt a ligament (grade III sprain)
(12 specimens)
(iv) Calcaneo- 1.06 ± 0.04 6.3 ± 0.5 (f) 0.3457 38 7.051E−02
fibular (16 spe- ±0.0552 (max) ±3 (f) ±0.690E−02
cimens)
(v) Posterior 0.82 ± 0.13 13.1 ± 1.6 (f) 0.2612 100 3.975E−02
talo-fibular ±0.0324 (max) ±15 (f) ±1.379E−02
(4 specimens)
(vi) Tibiotalar 0.80 ± 0.13 10.5 ± 1.1 (f) 0.7138 210 1.2282E−01
ligament ±0.0693 (max) ±23 (f) ±0.2504E−01
A PPENDIX B
(6 specimens)
S ECTION B.2
TABLE B.11
(Continued)

(f) Anterior cruciate

ligament
(WAINWRIGHT,
B IGGS and C URREY
[1979] (Reproduced

by permission of the
Princeton University
Press.)
(g) Anterior cruciate lig- force-elongation curves:

ament (W OO , P ETERSON
and O HLAND [1990]
(Reproduced by
permission of the
Orthopaedic Research
Society.)
385
386
TABLE B.11
(Continued)

(h) Mediacollateral
ligament
τmax = 4.0e–04–8.0e–04Gpa
E. Haug et al.
shear stress- force-strain curves:
deformation
(S HELTON , B UT-
LER and F EDER
[1993])
(Reproduced by
permission of
ASME International.)
(k) Human ACL in tension (H AUT [1993])

(16–26 years) 1.73 ± 0.27 (u) 44.3 ± 8.5 (f) 0.111 ± 0.026 σu = 37.8 ± 9.3 MPa (f)
(22–35 years) 2.16 ± 0.175 (u)
(48–86 years) 0.065 ± 0.024
A PPENDIX B
S ECTION B.2
TABLE B.11
(Continued)

(l) Spinal ligaments tensile response curves:

(PANJABI , J ORNEUS and
G REENSTEIN [1984])
(Reproduced by
permission of the
Orthopaedic Research
Society.)

(m) Human ankle/foot ligaments (PARENTEAU , V IANO and P ETIT [1996])
(i) force-strain curves force-strain curves:
(Reproduced by
permission of
ASME International.)
387
388
TABLE B.11
(Continued)

(ii) anterior talo-fibular 0.014–0.114 (y) 11–15 k = 0.006–0.023 kN/mm

ligament 0.064–0.286 (u) 24–44
(5 specimens)
(iii) calcaneo-fibular 0.053–0.259 (y) 21–39 k = 0.018-0.044 kN/mm
ligament 0.120–0.290 (u) 30–84
(6 specimens)
(iv) posterior talo-fibular 0.307 (y) 35 k = 0.037 kN/mm
ligament 0.307 (u) 35
(1 specimen)
E. Haug et al.
(v) deltoid 0.119–0.355 (y) 10–50 k = 0.026–0.068 kN/mm
ligament 0.239–0.507 (u) 45–79
(6 specimens)
(vi) talo-calcaneal 0.061–0.117 (y) 13–33 k = 0.021–0.031 kN/mm
ligament 0.078–0.156 (u) 13–33
(5 specimens)
(vii) plantar ligament 0.199–0.511 (y) 8–27 k = 0.039–0.495 kN/mm
(6 specimens) 0.238–0.506 (u) 12–32
(viii) metatarsal ligament 0.075–0.220 (y) 11–40 k = 0.016–0.038 kN/mm
(4 specimens) 0.103–0.247 (u) 0–54
(n) Human ankle/foot ligaments (N IGG , S KARVAN and F RANK [1990])

(i) anterior talo-fibular 0.067–0.193 6–60
ligament
(ii) calcaneo-fibular 0.265–0.327 27–81
ligament
A PPENDIX B
(iii) deltoid ligament 0.173–0.315 34–58
(f) = failure; (u) = ultimate; (m) = mean

tested as bone-ligament-bone test pieces, where the ligament and connected bones are
isolated and where the bones are subjected to relative displacements in the direction of
the fibers of the attached ligament. In the table, (f ) means “failure”, (u) means “ulti-
mate” and (m) means “mean”.
B.3. Brain mechanical properties
A synthesis of existing bibliography on the biomaterial behavior characterization of the

brain is presented in Tables B.12 and B.13. A list of variables is given after Table B.14.
B.4. Joint mechanical properties
The physiological motions of the synovial joints of the human skeleton can be mod-
eled approximately by computationally efficient point-like mechanical joint elements,
with six more or less constrained motion degrees of freedom. Motion ranges, stiffness
and resistance properties are found in the literature (ex: ROBBINS [1983], ROBBINS ,
S CHNEIDER , S NYDER , P FLUG and H AFFNER [1983]). Further data are given in Ta-
bles B.15 and B.16, in a coordinate system formed by the sagittal, coronal and horizontal
planes. The rotation movements of joints are performed about the longitudinal axis of
a body segment. The anatomically forbidden joint motions are often penalized by stiff
springs, while the natural motions are hardly penalized within the admissible motion
ranges and strongly penalized at the limits of these ranges. The corresponding stiffness
is also shown in Tables B.15 and B.16. Fig. B.1 shows typical non-linear moment-
rotation curves of the hip joint and of the subtalar joint of the ankle. The friction
coefficients of the synovial joints of the human skeleton vary between 0.005 and 0.04.
In cases when more detail about the local motion and response under load is required,
the joints can be modeled with their true articular surface geometry and their connecting
ligaments. Detailed finite element models are also applied in orthopedic analysis. The
gain in precision is obtained at a higher computational cost.
F IG . B.1. Typical moment-angle characteristics of the joints: (a) hip joint, and (b) ankle-subtalar joint (from
YANG and L ÖVSUND [1997] and PARENTEAU and V IANO [1996]). (Reproduced by permission of Chalmers
University of Technology.)
390
TABLE B.12
Brain material properties
Authors Experimental Characteristics
configuration
type and conditions ρ [kg/mm3 ] E [GPa] ν K [GPa] G [GPa] Other
D I M ASI , M ARCUS and E PPINGER linear visco elastic 6.895E−02 Gs = 3.4474E−05 β = 0.100 m s−1
[1991] Gl = 1.723E−05
E STES and M C E LHANEY [1970] incompressible with 2.070E+00
strain-rate independent
bulk modulus
(human, monkey)
E. Haug et al.
FALLENSTEIN , H ULCE and dynamic complex shear G1 = 6.00E−07 to
M ELVIN [1970] modulus by vibration tests 1.10E−06
at 10 Hz (human autopsy G2 = 3.00E−07 to
brain) 6.50E−07
F IROOZBAKSHK [1975] 1.000E−06 1.920E−04
G ALFORD and M C E LHANEY [1970] dynamic complex tensile E1 = 6.67E−05
modulus by vibration test E2 = 2.62E−05
at 34 Hz (human brain tis-
sue)
G OLDSMITH [1972] 7.800E−05 4.320E−10
L EE , M ELVIN and U ENO [1987] 1.000E−06 0.49 8.000E−05
M C E LHANEY, M ELVIN , ROBERTS dynamic complex shear G1 = 4.30E−07 G2/G1 = 0.72
and P ORTNOY [1973] modulus by vibration tests to 9.50E−07
at 9–10 Hz G2 = 3.50E−07
to 6.00E−07
A PPENDIX B
TABLE B.12
S ECTION B.4
(Continued)

configuration
M ARGULIES and T HIBAULT 1.060E−06 4.000E−05 0.5 2.000E+00 1, 380E−05

[1989]
O MMAYA [1967] 8.00E−06
−1.5E−05

ROSE and G ORDON [1974] 1.050E−06 2.100E+00
RUAN , K HALIL and K ING [1991] 1.040E−06 6.670E−05 0.48–0.499 2.190E+00 1.680E−03
RUAN , K HALIL and K ING [1994] linearly viscoelastic 1.279E−01 Gs = 5.28E−04 β = 0.035 m s−1
Gl = 1.68E−04
S HUGAR [1975] 6.720E−07 1.030E−05 0.5 2.100E+00 3.450E−06
S HUCK , H AYNES and F OGLE dynamic complex shear G1 = 8.30E−07 to

[1970] modulus by vibration tests 1.38E−04
at 2–400 Hz G2 = 3.40E−07 to
8.27E−05
S HUCK and A DVANI [1972] dynamic complex shear G1 = 7.60E−06 to

modulus by vibration tests 3.39E−05
at 5–350 Hz G2 = 2.76E−06 to
8.16E−05
391
392
TABLE B.12
(Continued)

configuration
T HIBAULT and M ARGULIES Age effect on complex G = 7.500E−07 to

[1996] shear modulus in 1.5938E−06
custom-designed G1 = 6.875E−07 to
oscillatory shear testing 1.1875E−06
device, at shear strain G2 = 1.875E−07 to
2.5% from 20–200 Hz, 1.00E−06
25 ◦ C, 100% humidity
E. Haug et al.
only in one location and
one direction (neonatal
(2–3 days) pigs: curve
data)
T ROSSEILLE , TARRIERE , 1.000E−06 2.400E−04 0.49–0.499
L AVASTE , G UILLON and
D OMONT [1992]
U ENO , M ELVIN , L I and 2.400E−04 0.49 4.000E−03 8.000E−05

L IGHTHALL [1995]
WARD and T HOMSON [1975] 1.040E−06 6.670E−05 0.48
WARD [1982] 6.500E−04 0.48–0.499

T URQUIER , K ANG , T ROSSEILLE , 1.140E−6 6.750E−04 0.48 5.625E−03 Gs = 5.28E−04 β = 0.035 m s−1
W ILLINGER , L AVASTE , Gl = 1.68E−04
A PPENDIX B
TARRIERE and D ÖMONT [1996]
S ECTION B.4
TABLE B.13
Brain material properties (white vs. gray matter and CSF)
configuration type ρ E ν K G Other
and conditions [kg/mm3 ] [GPa] [GPa] [GPa]
White matter:
TADA and N AGASHIMA [1994] 5.000E−05 0.49 perm. =
1.0E−11 mm2
poros. = 0.2

Z HOU, C. ET AL . [1996] 1.040E−06 0.4996 3.490E−01 2.680E−04
Z HOU, C. ET AL . [1996] S HUCK and A DVANI [1972] 2.190E−400 Gs = 4.10E−05 β = 0.70 ms−1
Gl = 7.6E−06
Gray matter:
1.0E−14 mm2
poros. = 0.2
Z HOU, C. ET AL . [1996] 1.040E−06 0.4996 2.190E−01 G = 1.68E−04
Z HOU, C. ET AL . [1996] S HUCK and A DVANI [1972] 2.190E+00 Gs = 3.40E−05 β = 0.70 ms−1
Gl = 6.3E−06
CSF (cerebro-spinal fluid):

RUAN , K HALIL and K ING [1991] 1.040E−06 0.489 2.190E−02 5.000E−04
RUAN , K HALIL and K ING [1991] 1.000E−06 2.190E+00
RUAN , K HALIL and K ING [1994] 1.040E−06 2.190E−02 5.000E−05
1.0E−05 mm2
poros. = 0.99
393
394
TABLE B.14
Brain material properties (other tissues)
Tissue & Experimental configuration Characteristics
authors type and conditions ρ E ν K G Other
[kg/mm3 ] [GPa] [GPa] [GPa]
Cerebellum, brainstem 1.040E−06 0.4996 2.19E−01 1.68E−04
RUAN , K HALIL and
K ING [1994]
Pia mater 1.133E−06 1.150E−02 0.45
Z HOU, Q. ET AL . [1996]
Bridging veins 1.133E−06 1.100E−04 0.45
Z HOU, Q. ET AL . [1996]
Dura, falx & tentorium 1.133E−06 3.150E−02 0.45
Z HOU, Q. ET AL . [1996]
E. Haug et al.
Dura/Falx tensile at frequency:
4.157E−02
M C E LHANEY, M ELVIN , 6.66E−05 msec−1 4.435E−02
ROBERTS and P ORTNOY 6.66E−04 msec−1 6.069E−02
[1973] 6.66E−03 msec−1
M ELVIN , M C E LHANEY 4.1382E−02 to

and ROBERTS [1970] 5.5176E−02
Membrane 1.133E−06 3.150E−02 0.45
RUAN , K HALIL and
K ING [1993]
Dura Mater square 15 mm on each side σue = 3.8E−02 ± 0.0018 [kg/mm2 ]
YAMADA [1970] expansive properties (curve e = 0.2σ e
σlim u
Fig. 198 p. 222)
(sample: cerebral rabbit)
rectangular strap 2 cm wide and T = 1.26 [kg/mm]
A PPENDIX B
2.5 cm long
shearing properties τu = 1.98 [kg/mm2 ]
(human adult average)
TABLE
List of variables used in Tables B.11 through B.14
ρ = density E1 = storage tensile modulus

E = Young’s modulus E2 = loss tensile modulus
ν = Poisson’s ratio β = decay factor
K = bulk modulus σue = ultimate expansive strength
G = shear modulus e = initial expansive strength
σlim
G1 = storage shear modulus T = shearing breaking load per unit width
G2 = loss shear modulus τu = ultimate shearing strength
Ge = equivalent shear modulus perm. = permeability
Gl = long term shear modulus poros. = porosity
Gs = short term shear modulus
The shear relaxation behavior is described by the time dependent shear modulus
G(t) = Gl + (Gs − Gl )e−βt , where t is time in a relaxation test.
B.5. Inter-vertebral joint mechanical properties
Fig. B.2 indicates the average range of motion for rotations between the vertebrae,
where C2–C7 are the vertebrae of the cerebral spine, T1–T12 of the thoracic spine,
L1–L5 of the lumbar spine and S1 indicates the sacrum. The range of rotational motion
between C1 (atlas) and C2 (axis) is about 35 degrees and accounts for about 50% of
the rotation of the head. The range of motion also differs between individuals, sexes
and is strongly age-dependent, decreasing by about 50% from youth to old age. (See
Table B.17.)
F IG . B.2. Inter-vertebral joint range of motion. (Reproduced by permission of Chalmers University of Tech-
nology.)
396
TABLE B.15
Upper limb joint mechanical properties
Joint Authors Experimental Characteristics
configuration Physiological Motion range Stiffness Mechanical
feature [deg] [Nm/deg] feature
Shoulder K APANDJI [1974a] Physiological motion Flexion 180 0–0.3 Ball joint
Extension 45–50 0–0.2
Abduction 140 0–0.3
Adduction 30–45 –
E. Haug et al.
F RANKEL and N ORDIN [1980] Static load Lateral rotation 80 0.3
Medial rotation 95 0.3
Elbow K APANDJI [1974a] Physiological motion Flexion 145–160 0–0.2 Pin joint
Extension 0 –
Abduction 0 –
Adduction 0 –
Lateral rotation 90 0.2
Wrist K APANDJI [1974a] Physiological motion Flexion 85 Cardan joint
Extension 85
Abduction 15
Adduction 45
A PPENDIX B
S ECTION B.5
TABLE B.16
Lower limb joint mechanical properties
configuration Physiological Motion range Stiffness Mechanical
feature [deg] [Nm/deg] feature
Hip K APANDJI [1974b] Physiological motion Flexion 90–145 0–2.5 Ball joint
Extension 20–30 1.2

F RANKEL and N ORDIN [1980] Abduction 45 0–1.2
M OW and H AYES [1991] Static load Adduction 30 0.8
Knee K APANDJI [1974b] Physiological motion Flexion 120–160 0–1.2 Pin joint
Extension 5–10 2.0
F RANKEL and N ORDIN [1980] Abduction 0 –
Adduction 0 –
Ankle K APANDJI [1974b] Physiological motion Tibia-Talar: Pin joint
Dorsiflexion 20–30 0.5
Plantarflexion 30–50 0.3
PARENTEAU and V IANO [1996] Quasi-static load Subtalar: Ball joint
Inversion 15–20 1
Eversion 10–15 1.5
Abduction 30–40
Adduction 25–35
397
398
TABLE B.17
Inter-vertebral joint mechanical properties
configuration Physiological Motion range Stiffness Other
feature [deg] [Nm/deg]
Lumbar K APANDJI [1974a] Physiological motion Flexion 60 1.0–2.1 Tolerance:
F RANKEL and N ORDIN [1980] Extension 35 0.3–1.8 Mflex = 145–185 Nm
Lateral flexion ±20 2.0
Static load Axial rotation ±5 0.9
Thoracic K APANDJI [1974a] Physiological motion Flexion 45 1.0–2.1 Tolerance:

F RANKEL and N ORDIN [1980] Extension 25 0.3–1.8 Mflex = 616 Nm
E. Haug et al.
Lateral flexion ±20 2.0 Mext = 240 Nm
Static load Axial rotation ±35 0.9
Cervical K APANDJI [1974a] Physiological motion Flexion 40 1.4

Extension 75 2.5
W ISMANS and S PENNY [1983] Dynamic load
lateral bending Lateral flexion ±45 0.4–2.2
torsion Axial rotation ±50 0.0–0.5
M C E LHANEY, D OHERTY, PAVER , Dynamic 10.5–67.5 [Nm/rad]
M YERS and G RAY [1988]
M ERTZ and PATRIC [1971] flexion–extension Tolerance:
Mflex = 56.7 Nm
Mext = 189 Nm
B OWMAN , S CHNEIDER , L USTAK , Torsion 2.71–3.74
A NDERSON and T HOMAS [1984]
A PPENDIX B
TABLE B.18
Articular cartilage and meniscus mechanical properties
Tissue Authors Experimental Characteristics

configuration Density Young’s Poisson’s Other material
[kg/m3 ] modulus ratio coefficients
[MPa] ν
Articular M OW and H AYES [1991] Tensile 0.3–1.0 0.1–0.4

Cartilage Solid matrix
YAMADA [1970] Compressive > 1.0
V IANO [1986] Instantaneous E = 12 0.42 G = 4.1 MPa
K = 2.5 MPa
Asymptotic E = 7.1 0.37 G = 2.6 MPa
K = 9.1 Mpa
A RMSTRONG , L AI 1000 E = 35 0.45
and M OW [1984]
Meniscus M OW and H AYES [1991] Tensile 0.1–0.6
B.6. Articular cartilage and meniscus mechanical properties
Cartilage is known to behave as a biphasic material, where a fluid seeps through a solid
porous matrix, which can lead to slow deformation under compressive loads. For the
short term behavior in impact studies these time and load dependent properties are
not dominant, and only the classical stiffness terms are required. To describe the com-
plete flow and deformational behavior of cartilage and meniscus the biphasic theory
was developed by Mow and coworkers. In this theory the solid matrix is linearly elas-
tic and isotropic, the solid matrix and interstitial fluid are intrinsically incompressible
and viscous dissipation is due to interstitial fluid flow relative to the solid matrix. (See
Table B.18.)
B.7. Inter-vertebral disc mechanical properties
The inter-vertebral discs assure the elastic coherence of the spinal column and they
provide a shock absorbing effect. Each disc is composed of the nucleus pulposus, the
annulus fibrosus, and a cartilageous end-plate. Between 70 to 90% of the nucleus pul-
posus by weight is water, and it takes up as much as 40 to 60% of the disc area. The
annulus fibrosus is a laminated and hence an anisotropic structure composed of several
concentric layers with fibers alternating at plus and minus 30 degree angles of inclina-
tion from the horizontal plane. The inner boundary of the annulus fibrosus is attached to
the cartilageous end-plate, and the outer surface is directly connected to vertebra body.
The discs play dominant role in sustaining the body against compressive load. Under
compression, the nucleus pulposus acts like a fluid in a cylinder made of the annulus
fibrosus. The discs show greater stiffness for the front/rearward inter-vertebral shear-
ing motion than for the side/side motion, with the annulus fibrosus rather the nucleus
pulposus making a major contribution.
The mechanical properties of the inter-vertebral discs are summarized in Table B.19.
Table B.20 summarizes age and region dependent tensile properties of the discs and
Tables B.21 and B.22 contain the average region dependent compressive and torsional
properties (S ONODA [1962]). Note that inter-vertebral discs from people in age group
20–39 have the greatest ultimate loads. Discs of females have a breaking moment about
five-sixths of that in males. The ultimate torsional strength and angle of twist in whole
discs of females are also less than in males.
B.8. Muscle mechanical properties
Skeletal muscles have active and passive properties, Fig. B.3. The active muscle action
is usually not of prime importance in car occupant impact simulations, except in low
energy collisions, where the activation of the muscles during bracing can modify the in-
jury pattern (example: rear impact/whiplash). In the following tables mostly the passive
mechanical behaviour of the skeletal muscles is documented.
A quasi-linear visco-elastic model was proposed to model the passive response of
skeletal muscle. Within the physiological muscle length, the passive muscle force is
usually much lower than the active force. At high elongation, nearing the physiological
limits of joint motion, the passive force increases rapidly and reaches the same level
and beyond as the maximum active force, while the active force drops to low values.
At high stretch, the axial muscle force is therefore dominated by the passive force.
At negative stretch velocities the active muscle force drops to almost zero at a given
reference velocity, while at positive stretch velocities this force will increase beyond
the activation level at zero stretch velocity. The Hill muscle model is often evoked in
simulation of the active and passive kinetics of skeletal muscles. This law is described
in Appendix C.
F IG . B.3. Muscle force-length and muscle force-velocity properties (Z AJAC [1989]). (Reproduced by per-
mission of Begell House, Inc.)
S ECTION B.8
TABLE B.19
Inter-vertebral disc mechanical properties
Tissue Authors Experimental Characteristics
configuration Density Young’s modulus Poisson ratio Shear modulus G12 Other
[MPa] ν12 [MPa]
Fibers G ALANTE [1967] Tensile 400–500
lumbar region E1 = 500 0.3 192
E2 = 500

Ground substance G OEL , M ONROE , G ILBERTSON and Compressive 2–4.2
B RINKMAN [1995] lumbar
U ENO and L IU [1987] Torsion E1 = 3 0.45 1
lumbar E2 = 3
Lamellae S KAGGS , W EIDENBAUM , I ATRIDIS , Tensile

R ATCLIFFE and M OW [1994] Posterior of disc E1 = 70 ± 42
Anterior of disc E1 = 106 ± 72
K ULAK , B ELYTSCHKO , S CHULTZ and E1 = 83 0.45 1.38
G ALANTE [1976] E2 = 2.07
Annulus fibrosus E BARA , I ATRIDIS , S ETTON , F OSTER , Tensile Ecirc = 5–50

M OW and W EIDENBAUM [1996] lumbar
Annulus fibrosus S PILKER , JAKOBS and S CHULTZ [1986] Modeling Ecirc = 33.4 0.5 0.189
Ez = 0.9
L IN , L IU , R AY and N IKRAVESH [1978] Ecirc = 22.4 0.45 3.92
Ez = 11.7
401
TABLE B.20
Age and region dependent tensile inter-vertebral disc properties
Region Characteristics
20–39 yr 40–79 yr Adult average
Fu σu εu Fu σu εu Fu σu εu
[kg] [kg/mm2 ] [%] [kg] [kg/mm2 ] [%] [kg] [kg/mm2 ] [%]
Cervical 105 ± 14.5 0.33 ± 0.02 89 ± 4.2 80 ± 8.6 0.29 ± 0.03 71 ± 3.6 88 0.30 77
Upper 142 ± 16.3 0.24 ± 0.01 55 ± 3.8 106 ± 9.4 0.20 ± 0.02 41 ± 2.1 118 0.21 46
thoracic
Lower 291 ± 21.5 0.26 ± 0.02 57 ± 6.3 220 ± 12.8 0.22 ± 0.01 40 ± 2.4 244 0.23 46
thoracic
Lumbar 394 ± 24.6 0.30 ± 0.01 68 ± 7.1 290 ± 19.5 0.24 ± 0.01 52 ± 6.2 325 0.26 59
Average 233 0.28 67 174 0.24 51 194 0.25 57
Ratio 1 1 1 0.75 0.85 0.76 0.83 0.89 0.85
Fu = breaking load in [kg]

σu = ultimate strength in [kg/mm2 ]
εu = ultimate elongation in [%]
TABLE B.21
Region dependent compressive inter-vertebral disc properties
Region Characteristics (40 to 59 years of age)

Breaking Load [kg] Ultimate Strength [kg/mm2 ] Ultimate Contraction [%]
Cervical 320 1.08 35.2

Upper thoracic 450 1.02 28.6
Lower thoracic 1150 1.08 31.4
Lumbar 1500 1.12 35.5
Average (855) 1.08 32.6
TABLE B.22
Region dependent torsional inter-vertebral disc properties
Region Characteristics (40 to 59 years of age)

Breaking Moment [kg cm] Ultimate Strength [kg/mm2 ] Ultimate Angle of Twist [deg]
Cervical 51 0.48 34
Upper thoracic 84 0.41 26
Middle thoracic 167 0.44 22
Lower thoracic 265 0.45 17
Lumbar 440 0.48 14
Average 201 0.45 23

TABLE B.23
Ultimate tensile strength of skeletal muscles (K ATAKE [1961])
Body segment Muscles Ultimate tensile strength [MPa]
Trunk Sternocleidomastoideus 0.19

Trapezius 0.16
Pectoralis major 0.13
Rectus abdominis 0.14
Upper extremity Biceps brachii 0.17
Triceps brachii 0.21
Flexor carpi radialis 0.15
Brachioradialis 0.18
Lower extremity Psoas major 0.12
Sartorius 0.30
Gracilis 0.20
Rectus femoris 0.10
Vastus medialis 0.15
Adductor longus 0.13
Semimembranous 0.13
Gastrocnemius 0.10
Tibialis anterior 0.22
F IG . B.4. Passive stress-strain curves in tension of skeletal muscles for 29 year old persons (YAMADA
[1970]). (Reproduced by permission of Lippincott, Williams and Wilkins.)
Table B.23 contains the ultimate tensile strength of a selection of skeletal muscles.
The corresponding ultimate passive muscle forces will be obtained by multiplying the
strength by the physiological cross section areas of each muscle. Fig. B.4 shows the non-
linear passive stress-strain curves of the muscles. Table B.24 contains age differences in
passive tensile properties of the rectus abdominis muscle. There is no significant sexual
difference in the ultimate strength of skeletal muscles. (See Table B.25.)
B.9. Tendon mechanical properties
Tendons connect muscles to the skeletal bones. There is almost no age effect on the
ultimate elongation, δu , a moderate age effect on the elastic modulus, E, and a marked
TABLE B.24
Age differences in passive tensile properties of rectus abdominis muscle (K ATAKE [1961])
Age group Adult

10–19 yr 20–29 yr 30–39 yr 40–49 yr 50–59 yr 60–69 yr 70–79 yr average
Ultimate tensile strength [g/mm2 = 0.01 Mpa]

19 ± 1.2 15 ± 0.6 13 ± 1.0 11 ± 0.6 10 ± 0.5 9 ± 0.3 9 ± 0.3 11
ratio 1.00 0.79 0.68 0.58 0.53 0.47 0.47
Ultimate percentage elongation

65 ± 1.2 64 ± 1.1 62 ± 0.7 61 ± 0.9 61 ± 1.5 58 ± 1.8 58 ± 1.8 61
ratio 1.00 0.98 0.95 0.94 0.94 0.89 0.89
TABLE B.25
Summary of studies on mechanical properties of muscles
Authors Experimental configuration Characteristics

σu [MPa] Other
YAMADA [1970] Experiments on various ani- 0.1–0.32

mal and human specimens
W INTERS and S TARK [1985], 0.5–1.0
W INTERS and S TARK [1988]
S CHNECK [1992] C = damping coefficient 0.2–1.0 C = 10–1000 Ns/m
K = stiffness of whole
muscle K = 32.5–250 kN/m
age effect on the ultimate tensile strength, σu . The ultimate tensile strength can exceed
the insertion strength of a tendon. There is a significant strain rate effect on the elastic
modulus. The longitudinal tendon strips of the supraspinatus muscle are not of equal
strength. (See Table B.26.)
B.10. Skin mechanical properties
Skin is a non-linear elastic material and its response is orthotropic. In human models,
skin should be modeled as the enveloping membrane of flesh and fatty tissue. Skin
assures the stability of the female breasts, which consists of fatty tissue, enveloped by
resistant skin. The behavior of skin is best modeled as a material with perpendicular
layers of fibers, each described with non-linear stress-elongation curves (see Figs. B.5–
B.7). Some mechanical properties are listed in Table B.27.
B.11. Internal organ mechanical properties
Rather little information is available on the mechanical properties of the internal organs.
Table B.28 contains some preliminary data on the lungs, liver and spleen. The data on
TABLE B.26
Mechanical properties of tendons
Type of tissue Authors Experimental E [GPa] ν σu [MPa] δu

configuration Poissons Ultimate Ultimate
type and ratio strength elongation
conditions
M OW and 1.2–1.8 Incomp.
H AYES [1991]
Achilles L EWIS and Tension (fiber 2.00 ± 0.99 0.4

tendon S HAW [1997a], direction)
L EWIS and 0.1sec−1 and
S HAW [1997b] 1 sec−1
embalmed, age
36–100
Patellar tendon J OHNSON , Tension (fiber

T RAMAGLINI , direction) fresh
L EVINE , O NO , frozen,
C HOI and W OO non-irradiated
[1994] (i) age 29–50 0.66 ± 0.266 64.7 ± 15 0.14 ± 0.06
(ii) age 64–93 0.504 ± 0.222 53.6 ± 10 0.15 ± 0.05
Human tendon V OIGT, B OJSEN - Tension (fiber 1.2 50 0.06

M OLLER , direction)
S IMONSEN and (yield point)
DYHRE -
P OULSEN
[1995]
Patellar tendon W OO , J OHNSON Tension (fiber 0.58

and S MITH direction)
[1993] Stress–Relaxation
Finger flexor P RING , A MIS Tension (fiber 0.13

and C OOMBS direction)
[1985]
Supraspinatus I TOI , Tension (fiber

B ERGLUND , direction) fresh
G RABOWSKI , (i) anterior 16.5 ± 7.1
S CHULTZ , (ii) middle 6.0 ± 2.6
G ROWNEY, (iii) posterior 4.1 ± 1.3
M ORREY and
A N [1995]
the liver and spleen are estimations used in a project on virtual abdominal surgery, DAN
and M ILCENT [2002].
Fig. B.8 shows the non-linear force-displacement curve of an entire human liver un-
der compression between two parallel plates under quasi-static loading, DAN [1995].
The tested liver was pressurized in the sense that the natural in- and outflow of body
F IG . B.5. Stress-strain curves in tension of the skin of persons 20 to 29 years of age (YAMADA [1970]).
(Reproduced by permission of Lippincott, Williams and Wilkins.)
F IG . B.6. Force-stretch relation of rabbit skin at stretch rate 0.2 mm/s (L ANIR and F UNG [1974]). (Repro-
duced by permission of the Journal of Biomechanics.)
F IG . B.7. Stress-strain relation of rat skin (O XLUND and A NDREASSON [1980]). (Reproduced by permission
of Blackwell Publishing Ltd.)
fluids was maintained artificially with a circulating substitute fluid. The liver as a whole
body was then loaded. It resists the applied compression load with the combined action
of the stored circulating fluid and the resistance of its bulk material.
TABLE B.27
Mechanical properties of skin
Tissue Authors Experimental E [Gpa] Tensile σu δmax

Configuration type breaking [UTS] Ultimate
and conditions load per [kg/mm2 ] percentage
unit elongation
width
[kg/mm2 ]
Skin of M ANSCHOT and Tension

the calf B RAKKEE [1986a], Slow rates in Vivo
M ANSCHOT and (i) along tibial axis Non-linear
B RAKKEE [1986b] (ii) across tibial Non-linear
axis
Skin YAMADA [1970] Tension human
skin
(i) 30–49 years Non-linear 0.2 to 3.4 0.29 to 1.47 –
(ii) 10–29 years Non-linear – – 56 to 144
(iii) Adult average Non-linear 0.2 to 3.1 0.26 to 1.32 43 to 111
TABLE B.28
Mechanical properties of some organs
Tissue Authors Experimental Density ρ Modulus Poisson Shear mod.

configuration [kg/m3 ] E [Mpa] Ratio ν G [Mpa]
type and
conditions
Lungs H OPPIN , L EE Compression E/Pt∗ = 4 0.3

and D AWSON (3-D loading)
[1975]
H AJJI , W ILSON G/07Pt = 1 to 1.5
and L AI -F OOK
[1979]
Liver D AN [1999], Small test 1158 0.3E–4 to 5E–4 0.4

D AN and cylinders of
M ILCENT [2002] parenchyma
Venous vessels 1168 0.158 0.49
Glisson capsule 1168 1.0 0.49
Spleen C ARTER [1999] (Indentation tests (much smaller)
for the study of
force feedback in
virtual surgery)
* Pt = the transpulmonary pressure
Fig. B.9 represents static compression results of small test cylinders of pure liver
parenchyma (diameter 12 mm by 15 mm initial height), from which the interstitial fluid
can escape, DAN [1999]. It indicates a non-linear distribution of the elastic modulus,
F IG . B.8. Whole human liver force vs. displacement curves (D AN [1995]). (Private communication by the
author.)
F IG . B.9. Liver parenchyma elastic modulus vs. displacement curves (D AN [1999]). (Private communication
by the author.)
F IG . B.10. Indentation tests on the liver and spleen of pigs (C ARTER [1999]). (Reproduced by permission
of Fiona Carter of the University of Dundee.)
and the response of the so isolated material is much weaker than the response of the
pressurized complete organ.
Fig. B.10 shows results of indentation tests on pig livers and spleens, C ARTER [1999],
which were performed in projects on the study of haptic force feedback in virtual
surgery. The picture shows that the spleen is much less resistant than the liver.
F IG . B.11. Typical drop test lesions in porcine kidneys (B SCHLEIPFER [2002]). (Reproduced by permission
of Dr. med. Thomas Bschleipfer.)
FAZEKAS , KOSA , J OBBA and M ESZARO [1971a], FAZEKAS , KOSA , J OBBA and
M ESZARO [1971b], FAZEKAS , KOSA , J OBBA and M ESZARO [1972] (in German) pub-
lished data on the compression resistance of the whole human cadaver liver, kidneys
and spleen, respectively. They found that the liver showed superficial ruptures at a
compressive stress of 169 kPa and multiple ruptures at 320 kPa. The first superficial
ruptures of the spleen occurred at a compressive stress of 44 kPa and of the kidneys
at a load of 60.2 ± 28.2 daN, the latter of which showed multiple ruptures at loads of
109.44 ± 51.4 daN.
BAUDER [1985] (in German) investigated the compressive resistance of the isolated
human liver with blunt drop weight impact tests. The tests showed for 3–4 m/s impact
velocities mean compressive loads of the organ of 175.6 ± 39.2 daN, which were asso-
ciated with mean compressive deformations of 29.5 ± 3.5 mm. The observed injuries
were contusions, superficial ruptures and crushing of the livers. The thickness of the
organs and the portion of connective tissue were important parameters for the severity
of the injuries.
B SCHLEIPFER [2002] investigated the lesions inflicted on isolated porcine kidneys
under blunt drop test impacts, from heights of 0.1 to 1.0 m, with a cylindrical impactor
(∅10 cm) of mass 1.45 kg (1.4 to 14.2 Joule), with and without ligatured urethers.
Typical lesions found in the tests are shown in Fig. B.11, for the first lesions at low
load, (a), for severe lesions at 0.8 m drop height, views, (b), and sections through the
organ, (c).
Fig. B.12, finally, gives an overview on the nonlinear visco-elastic modulus response
(Pa) of live porcine livers, based on in vivo semi-infinite elastic body indentation with
a vibrating cylindrical indenter, as a function of vibration frequency (Hz) and median
F IG . B.12. In vivo elastic modulus measurements of live porcine livers (O TTENSMEYER and S ALISBURY
[2001]). (Reproduced by permission of Springer Verlag.)
applied normal stress (Pa) (after OTTENSMEYER and S ALISBURY [2001]). This work
was performed for calibrating instrument force feedback in virtual surgery.
Summary Table B.29 is taken from YAMADA [1970]. It contains average ultimate
strength (stress) and percentage elongation (strain) for a variety of human tissues and
organs.
Data about many internal organs of humans are still missing. Table B.30 lists some
such data found from animals (YAMADA [1970]).
Further data on liver and kidneys of rhesus monkeys are reported by M ELVIN , S TAL -
NAKER and ROBERTS [1973]. These data are used by M ILLER [2000] in modelling
constitutive relationships of the abdominal organs.
B.12. Quasi-linear visco-elastic solids
Many biological tissues exhibit visco-elastic behaviour. A simple such law is described
next. The deviatoric response of a linear viscoelastic solid material is governed by
Zener’s model, which can be considered as a Maxwell spring-dashpot model in par-
allel with a spring, Fig. B.13.
The “slow motion” response to small strain rates, dε/dt, is governed by the long term
shear modulus, G∞ (≡ Gl ), while the instantaneous response to a step loading, H (t), is
according to the long term modulus G0 (≡ Gs ).
The elastic behaviour of this material is described by the deformation rate dependent
shear modulus, G, and by the constant bulk modulus, K.
E
Shear modulus G = where E = Young’s Modulus
2(1 + ν)
and
E
Bulk modulus K = ν = Poisson’s Ratio.
3(1 − 2ν)
TABLE B.29
S ECTION B.12
Average adult human mechanical tissue properties (YAMADA [1970])
Tissue σu [kg/mm2 ] δmax [%] Tissue σu [kg/mm2 ] δmax [%]
Hair 19.7 40 Ureter (L) 0.18 36
Compact bone (femur) 10.9 1.4 Mixed arterial tissue (L) 0.17 87
Chorda tendinea 6.4 33 Venous tissue (L) 0.17 89
Tendinous tissue (calcaneal) 5.4 9 Umbilical cord (mature fetus) 0.15 59
Nail 1.8 14 Mixed arterial tissue (T) 0.14 69
Fascia 1.4 16 Muscular arterial tissue (L) 0.14 102
Nerve (secondary fiber bundle) 1.3 18 Spinal dura mater (T) 0.13 34
Fibrocartilage (annulus fibrosus) (L) 1.3 14 Spongy bone (vertebra) 0.12 0.6
Skin (thorax, neck) 1.3 90 Coronary artery (L) 0.11 64

Spinal dura mater (L) 1.1 21 Renal calyx (L) 0.11 35
Skin (abdomen, back, foot, arm) 0.97 90 Elastic arterial tissue (T) 0.10 82
Skin (leg, hand) 0.74 90 Muscular arterial tissue (T) 0.10 75
Sclera (E) 0.69 17 Cardiac valve (R) 0.094 17
Fibrocartilage (annulus fibrosus) (T) 0.53 12 Elastic arterial tissue (L) 0.08 80
Sclera (M) 0.48 17 Large intestine (L) 0.069 117
Skin (face, head, genitals) 0.38 69 Esophagus (L) 0.06 73
Vertebra 0.35 0.8 Stomach (L) 0.056 93
Cornea 0.35 15 Small intestine (L) 0.056 43
Auricle 0.34 26 Small intestine (T) 0.053 89
Elastic cartilage (auricle) 0.31 26 Renal calyx (T) 0.048 48
Thyroid cartilage (L) 0.30 15 Large intestine (T) 0.045 137
Venous tissue (T) 0.30 66 Ureter (T) 0.045 89
Hyaline cartilage (costal) 0.29 18 Stomach (T) 0.044 127
Intervertebral disc 0.28 57 Tracheal membranous wall (T) 0.036 81
Cardiac valve (C) 0.25 13 Urinary bladder 0.024 126
Tracheal cartilage 0.24 18 Papillary muscle tissue 0.023 30
Amnion (normal labor) 0.24 42 Esophagus (T) 0.018 124
Renal fibrous capsule 0.23 29 Skeletal muscle tissue (rectus abdominis) 0.011 61
Tracheal membrane wall (L) 0.22 61 Cardiac muscle tissue 0.011 64
Tracheal intercartlagnious membrane 0.19 138 Renal parynchyma 0.005 52
411
(C) = circumferentially, (E) = equatorially, (L) = longitudinally, (M) = meridionally, (R) = radially, (T) = transversely.
TABLE B.30
Mechanical properties of some internal organs of animals (YAMADA [1970])
Tissue Animal σu δmax

[kg/mm2 ] [%]
Liver parenchyma Rabbits 0.0024 46

Gall bladder Rabbits 0.21 53
Uterus Rabbits 0.018 150
Cerebral dura mater Rabbits 0.038
F IG . B.13. Zener type model.
The shear relaxation behaviour is given by the response to a step function, and is
described by the shear relaxation modulus

G(t) = G∞ + (G0 − G∞ )e−βt ≡Gl + (Gs − Gl )e−βt .

Time t is the current simulation time, but relaxation starts only when the material expe-
riences a deviatoric strain. The decay constant, β, has the unit of (time)−1 , which must
be consistent with the chosen time units.
The deviatoric stress rate, sij , depends on the shear relaxation modulus as follows
t
sij = 2 G(t − τ )Dij (τ ) dτ,
τ =0
where Dij is the deviatoric velocity strain tensor. The above convolution expression for
the deviatoric stress rate may be understood as follows: the deviatoric strain is approx-
imated by a piecewise constant function. The material responds to each step function
H (τ ) following the relaxation law. This material model does not describe volumetric
viscous effects, as might be present in the compression of foams.
B.13. Further references on biomaterials
The literature on biomaterials and related subjects is abundant. Some references are
indicated in this appendix. While many of the indicated references deal with the ex-
perimental evaluation of biomaterial properties, others deal with the aspects of their
modeling, the use of these materials in biomechanical models and the characterization
of trauma and injury.
Many further references for biomaterials could be cited by separate topics on bones,
ligaments, brain, joints, spine and inter-vertebral discs, muscles, tendons, skin and or-
gans. These references are not mentioned explicitly in this appendix, but constitute
further valuable sources of historical and actual information.
A PPENDIX C
The Hill Muscle Model

Skeletal muscles. The Hill muscle model is one of the simplest phenomenological en-
gineering models of the active and passive biomechanical behaviour of skeletal muscles
(H ILL [1970]). Its simplest implementation is with bar finite elements. More involved
implementations can be in 2D and 3D composite finite elements, where the composite
fibers are assigned the properties of Hill-type muscle models.
Fig. C.1 gives an overview on the anatomical detail of skeletal muscles.
Active voluntary muscle contraction can be considered a material behaviour that has
no parallel in conventional engineering material models. Whereas the passive mechan-
ical impact behaviour of biological tissues can often be approximated using standard
engineering material models, active muscle behaviour clearly distinguishes living and
non-living materials. For this reason it is interesting to briefly outline the standard Hill
muscle model.
Standard Hill muscle model. This model, its implementation into a crash code and its
application is described by W ITTEK and K AJZER [1995], W ITTEK and K AJZER [1997];
W ITTEK , H AUG and K AJZER [1999]; W ITTEK , K AJZER and H AUG [1999]; W ITTEK ,
O NO and K AJZER [1999]; W ITTEK , O NO , K AJZER , Ö RTENGREN and I NAMI [2001].
Authors K AJZER , Z HOU , K HALIL and K ING [1996] describe the application of mod-
eling of ligaments and muscles under transient loads.
Fig. C.2 summarizes the Hill muscle bar model.
Inset (a) of Fig. C.2 is an overview of the types of skeletal muscles (W IRHED
[1985]). Inset (b) shows the schematics of the Hill model for a fusiform tendon-muscle-
tendon assembly. Inset (c) depicts the (normalized) active muscle component force
versus length diagrams, FCE /Fmax . Inset (d) contains the active component force ver-
sus (normalized) stretch velocity diagram, FV (V /Vmax), and inset (e) shows the active
component activation versus time function, Na (t).
Inset (b) of Fig. C.2 shows a simple mechanical model of a fusiform muscle with
the contractile sub-element, (CE), the parallel elastic sub-element, (PE) and the parallel
dashpot sub-element, (DE), of its active “muscle” element, and the nonlinear spring sub-
element, (SE), and dashpot sub-element, (DSE), of its two “tendon” elements, switched
in series. The tendons are not discussed here because as passive materials their mechan-
ical response can be approximated with standard engineering visco-elastic-damaging
material models. Their action can be modeled by arranging serial bars together with the
central muscle bar element.
413
414 E. Haug et al. A PPENDIX C
F IG . C.1. Skeletal muscle structure. (Reproduced by permission of the Longman Group UK Ltd.)
For the central “muscle” bar element, the total muscle force consists in an active and
a passive component,
Fmuscle = Factive muscle + Fpassive muscle = FCE + (FPE + FDE ). (C.1)
The normalized “active branch” of the muscle force, FCE /Fmax acting in the contrac-
tile sub-element, CE, is modeled by
FCE /Fmax = Na (t)FV (V /Vmax)FL (L/Lopt ) = Factive muscle/Fmax . (C.2)
In this expression Fmax = σ Aphys is the maximum muscle force at 100% voluntary
muscle activation, with the maximum active muscle stress σ ∼ = 0.001 Gpa, which is
fairly intrinsic to all skeletal muscles, and Aphys = physiological cross section area
The Hill muscle model 415
F IG . C.2. Hill’s model of skeletal muscles. (Insets (a), (b) anatomical drawings: Reproduced by permission
of Rolf Wirhed, W IRHED [1985].)
of the muscle; Na (t) is the neurological muscle activation state versus time function
(voluntary and reflexes); FV (V /Vmax ) is a muscle stretch velocity dependent function,
where Vmax is a reference muscle stretch velocity; V = dL/dt is the muscle length
rate of change or stretch velocity; FL (L/Lopt ) is a muscle length dependent shape
function, where L is the current length and Lopt is the optimal length of the skeletal
muscle “at rest”, at which the voluntary muscle force can reach its peaks. The opti-
mal muscle length is sometimes attributed to the freely floating position of a dormant
astronaut.
Inset (c) of Fig. C.2 shows for the contractile sub-element (CE) the stationary (Fv = 1
at V = 0) muscle force-elongation curves, FCE /Fmax , activated at Na = 25, 50 and
100%, over a normalized length range of L/Lopt between about 0.5 and 1.5 with the
length dependency curve, FL (L/Lopt ) (thin lines). Other length shape functions are
possible, depending on a shape factor, Csh , according to

FL (L/Lopt ) = exp −(L/Lopt − 1)/Csh . (C.3)

Inset (c) also shows the normalized “passive branch” contributed as the parallel elastic
sub-element (PE) force-length response, which is due to the cohesive material resistance
when the active fibers are not activated. The normalized passive forces, FPE /Fmax , can
be calculated from

FPE /Fmax = 1/ exp(CPE ) − 1 exp (CPE /PEmax ) (L/Lfib ) − 1 − 1 . (C.4)

In this expression CPE is a shape parameter of the passive force-length curve,
PEmax = L/Lfib at is the muscle stretch when the passive force FPE reaches the value
of Fmax and Lfib is a characteristic fiber length, often set to Lfib = Lopt .
If the muscle elongates at a stretch velocity of V = dL/dt, then the parallel passive
dashpot element (DE) responds with the force
FDE = CDE V , (C.5)
where CDE is the damping coefficient of the assumed linear parallel dashpot element,
DE.
The curves drawn with thick lines in inset (c) schematize the superimposed active and
passive muscle forces under isometric conditions, i.e., when the shortening or lengthen-
ing stretch velocities of the muscle are small or zero, V ≈ 0,
Fmuscle = FCE + FPE + (FDE = 0) = Factive muscle + Fpassive muscle . (C.6)
Inset (d) of Fig. C.2 presents the velocity dependent function, FV (V /Vmax), of the
active muscle force FCE . This function can be interpreted in classical engineering
terms as a nonlinear dashpot, as opposed to the familiar linear dashpot force-velocity
curve shown for comparison in the diagram. The curves can be constructed from three
branches as follows.

0 for v = V /Vmax −1,
FV (V /Vmax ) = (1 + v)/(1 − v/C short ) for −1 < v 0, (C.7)
 (1 + vC /C )/(1 + v/C ) for v > 0,
mvl leng leng
where Cshort is a shape parameter for the non-zero curve segment at shortening stretch
velocities, Cleng is a shape parameter for the curve segment at lengthening stretch ve-
locities, Cmvl is the asymptotic value of the curve for large positive stretch velocities,
v = V /Vmax 1.
The physiological origin of the stretch velocity dependency seems to stem from the
actions of the so-called cross-bridges inside the actine-myosine components of the sar-
comere cells of the active muscle fibers, see the bottom zooms of Fig. C.1 and insets (c)
and (d) of Fig. C.2. At zero stretch velocity, V = 0, the muscle can afford an isometric
active muscle force, when the cross bridges of the recruited muscle fibers continually
connect, flex forward and disconnect the telescoping actine and myosin muscle fiber
components. The number of recruited fibers can be modeled with the percentage level
of the (normalized) activation function, 0 Na (t) 1. The continuous process of con-
nection, flexion and disconnection creates a forward motion of the myosin fibers into
the actine tubes, which counteracts the backward slipping motion due to the constant
pull of the section force. This action can be compared roughly to the action of rowers
in a boat in still waters, who must keep rowing on the spot in order to create a steady
pull on a rope that retains their boat in place. If the retaining rope is cut, the boat will
move forward, which corresponds to the shortening of the unconstrained muscle if the
external force vanishes. The image of the tied rowers also helps understanding how
physiological energy must be spent in order to keep a muscle at the same length un-
der active tension, i.e., when the muscle does no external mechanical work. Although
the tied down boat does not move, the rowers will fatigue and eventually stop rowing.
Furthermore, it is easy to understand why in inset (c) the isometric active force-length
curves, FL (L/Lopt ), are not constant with the muscle length. If the muscle is longer
than optimal, L > Lopt , then the overlap of the myosin and actine components in a sar-
comere decreases in length, and less connecting cross bridges are available to create
the active muscle force. On the other hand, if the muscle has shortened, L < Lopt , the
efficiency of the cross bridge action decreases because of the hindrance created by the
shortening.
The shape of the velocity dependent function, FV (V /Vmax), in inset (d) of Fig. C.2 is
discussed next. If to a muscle at a given instantaneous length, L, and under a given active
force, F , a positive stretch velocity is imposed, V /Vmax > 0, the connected bridges tend
to be pulled in fiber direction and the force output at the same voluntary activation level
increases by the factor FV (V /Vmax ) > 1. At negative stretch velocities the connecting
bridges do not re-connect fast enough to make up for the negative length rate of change
of the muscle fibers, and the force output falls drastically and reaches the value of zero
at negative stretch velocity V = −Vmax . This again might be compared to rowers in a
boat who are more efficient when rowing downstream than upstream. When the face
stream velocity becomes equal to the rowing velocity, the action of the rowers will no
longer produce any force on the retaining rope.
Inset (e) of Fig. C.2 shows the active muscle state function Na (t). In a simplified
approach, this function depends on a muscle neurological reflex time, Treflex , which
is the time that elapses between, say, the onset of an impact event where a defensive
muscle action should ideally start and the time when the activation actually starts. After
the reflex time has elapsed, a neuro-control flag, u(t), is set equal to one, and the muscle
activation process sets in,

0 for t Treflex ,
u(t) = (C.8)
1 for t > Treflex .
Reflex times for skeletal muscles are known to range from about 25 to 100 milliseconds
(ms) and the reflex time is set to 80 ms in inset (e). After the reflex time has passed,
the muscle force must be activated. This physiological process takes a certain time,
and the maximum muscle force occurs at about 250 ms in inset (e). More details about
skeletal muscle excitation and activation can be found in the literature cited in the given
references.
The muscle activation function of the Hill model can be calculated from the differen-
tial equations

dNe (t)/dt = u(t) − Ne (t) /Tne ,

dNa (t)/dt = Ne (t) − Na (t) /Ta , (C.9)

where Ne (t) is the neuro-muscular excitation function, Na (t) is the muscle force exci-
tation function, 0.02 Tne 0.05 s and 0.005 Ta 0.02 s are time constants. For the
assumed binary form of the neural control flag u(t), there exists an analytical solution
of the form

Ainit for t Treflex ,
Na(t) = (C.10)
1 + Aa /Ba + Ane /Bne for t > Treflex ,
where
Aa = (Ainit − 1)(Ta − Tne ) − Tne ,

Ba = (Ta − Tne ) exp (t − Treflex )/Ta ,

Ane = Tne ,

Bne = (Ta − Tne ) exp (t − Treflex )/Tne ,

Ainit = Amin = 0.005,
Tne = C1 + C2 mCslow ,
Ta = B1 + B2 m(Cslow )2 ,
C1 = 0.025 s,
C2 = 0.01 s,
B1 = 0.005 s,
B2 = 0.0005 s,
m = muscle mass in grams,
Cslow = fraction of slow muscle fibers.
Applications. The mechanical behaviour of the skeletal muscles in the directions of

their fibers can be modeled to first order accuracy by Hill-type muscle bars. Each Hill-
type muscle bar element is characterized by the physiological cross section area of the
muscle, cut perpendicular to the fibers, and by the muscle fiber stretch and stretch ve-
locity dependent active and passive mechanical properties of the Hill muscle model,
described above. The bars cannot, in general, transmit compressive forces.
Fig. C.3 shows a couple of postures and the maximum sustainable limit loads as cal-
culated from the muscled skeleton model, compared to values found in the literature
(B OUISSET and M ATON [1995], p. 135). The calculated values follow from the appli-
cation of the optimization process described in Chapter II, Section 6, where the applied
loads were incremented until the process could no longer find a solution for the given
F IG . C.3. Validation of the limit load muscle force calculation (B ENES [2002]).
posture, i.e., the respective limit loads were attained. The shown model has a total of
196 anatomical muscle groups, modeled by 668 fasciae (bars). The colors of the bars
range between red (100% activation) and blue (no activation). The comparison with
the reference values is considered fair in view of the sensitivity of the results to the
joint geometry, the lever arms of the muscle bars, the muscle section area and trajec-
tory, the true strategy of activation, the chosen objective function (physiological muscle
energy/work) in the optimization process and the uncertainty of the experimental con-
ditions, such as the exact anatomy of the volunteers, the exact posture and the point of
load application (B ENES [2002]).
References on muscle materials. The following references deal with the structure and
the modeling of skeletal muscle: BAHLER , FALES and Z IERLER [1968] on the dynamic
properties of skeletal muscle; C OLE , B OGERT, H ERZOG and G ERRITSEN [1996] on
modeling of forces in stretched muscles; C RAWFORD AND JAMES [1980] on the design
of muscles; H ARRY, WARD , H EGLUD , M ORGAN and M C M AHON [1990] on cross
bridge action; H ERZOG [1994] on the biomechanics of the musculo-skeletal system;
H AWKINS and B EY [1994], H AWKINS and B EY [1997] on muscle-tendon mechanics
and mechanical properties; H ILL [1970] on experiments in muscle mechanics; K IRSCH ,
B OSKOV and RYMER [1994] on stiffness of moving cat muscles; K RYLOW and S ANDE -
ROCK [1996] on dynamic force response of muscles under excentric contraction; M A
and Z AHALAK [1991] on a distribution-moment model of energetics in skeletal mus-
cle; M ORGAN [1990] on the behaviour of muscle under active lengthening; M YERS ,
VAN E E , C AMACHO , W OOLLEY and B EST [1995] on the structural properties of mam-
malian skeletal muscle in the neck; R ACK and W ESTBURY [1969] on the effect of length
rate on tension in muscles; S CHNECK [1992] on the mechanics of muscle; VANCE ,
S OLOMONOV, BARATTA , Z EMBO and D’A MBROSIA [1994] on the comparison of two
muscle models; WANK and G UTEWORT [1993] on the simulation of muscular contrac-
tion with regard to physiological parameters; W INTERS and S TARK [1985], W INTERS
and S TARK [1988] on muscle modeling and mechanical properties; Z AJAC [1989] on
muscle and tendon properties and models; Z UURBIER , E VERARD , VAN D ER W EES
and H UIJING [1994] on the force-length characteristics in active and passive muscles.
A PPENDIX D
Airbag Models
The following paragraphs are based mainly on Pam-Safe documentation of ESI Soft-
ware.
Airbags. Airbags are considered volumes of ideal gas that are enclosed by a flexible
envelope, Fig. D.1. Their physics requires fluid-structure interaction (FSI) simulation.
In simple airbag models, pressure and temperature of the gas are assumed to be distrib-
uted uniformly throughout the airbag volume. More complex approaches (not described
here) model the enclosed gas and the envelope separately, using fluid-solid and multi-
physics formulations. The gas can be confined by a bag made of flexible (visco-)elastic
industrial fabric or of any other material, such as the walls of hollow organs.
Airbag models can simulate gas inflow and outflow through orifices, and leakage
of gas through the fabric of the envelope can be defined. The input data for the gas
are atmospheric pressure p, temperature T , constants γ and R, where R is the perfect
gas constant from pV = nRT , where n is the number of moles in the volume. In the
described simple airbag model, the gas is confined in a single chamber and it obeys
the thermodynamic equation of a perfect gas. Solution of that equation at each explicit
structural solution time step yields a pressure load to be applied to the inside of the en-
velope. Solution of the equations of motion of the pressurized envelope yields a volume
change to be applied to the enclosed gas. This process is repeated over the duration of
the simulation.
Airbag gas model. The volume of enclosed gas is subject to the equation of state
pV γ = constant, where p is the pressure, V is the volume and γ = cp /cv is the specific
heat ratio for a perfect gas under adiabatic conditions, and where cp and cv are the
specific heat of the gas at constant pressure and at constant volume, respectively. The
gas constant is defined as
R = cp − cv = Ru /W,
where Ru is the universal gas constant related to moles and W is the molecular weight
of the gas.
For nitrogen gas, N2 , one has in SI-units (m, kg, s) and with the molecular weight
W (N2 ) = 0.028014 kg/mole
421
422 E. Haug et al. A PPENDIX D
F IG . D.1. Basic airbag model (schematic) (ESI Software).
F IG . D.2. Basic airbag model for the lungs (ESI software).
cp = 1038 J/(kgK), cp = 29.08 J/(mole K),

cv = 741 J/(kgK), cv = 20.76 J/(mole K),
R = 297 J/(kgK), Ru = 8.32 J/(mole K),
where J = Joules = Nm (Newton meters), K = ◦ Kelvin and cp and cv are the molar
heat capacities.
The gauge pressure is the differential pressure between the airbag pressure and the
atmospheric pressure, Fig. D.1,
pgauge = p1 − pa .
The specific heat ratio, γ , is defined as
γ = cp /cv = cp /(cp − R),

which for nitrogen gas (N2 ) is equal to 1.4. The meaning of the remaining parameters
is illustrated on Fig. D.1.
Bio-bag models. Bio-bag models are derived from airbag models to model hollow
organs. Hollow organs have flexible walls and are filled with quasi-incompressible ma-
terial (blood, body fluids, food), but also (partly or fully) with compressible material
(gas in stomach or lungs). Airbag gas models can be used to approximate quasi-
incompressible fluid-filled hollow organs by re-setting their input data to model a linear
Airbag models 423
F IG . D.3. Bio-bag models of the mediastinum and heart (ESI software).
pressure-volume equation of the type
p = p0 + K(ρ/ρ0 − 1),
where K is the desired bulk modulus for modeling quasi-incompressibility, p0 is the

initial pressure and ρ and ρ0 are the mass density in the compressed and uncompressed
gas, respectively. This can be achieved approximately by setting cp to a large value
(isothermal conditions) with γ = cp /cv close to 1.
424 E. Haug et al. A PPENDIX D
F IG . D.4. Alternative airbag modeling techniques (ESI software).
Examples. The following examples demonstrate the use of “bio-bags” in the modeling
of the hollow internal organs.
Fig. D.2 is an approximate mechanical model of the lungs, which uses the basic
airbag model with perfect gas properties. A fictitious envelope encloses a space filled
with air, and vent holes can be provided to simulate the expulsion of air from a violent
compression of the thorax in an impact.
This model is relatively efficient when the lungs must not be simulated in detail, but
when only their resistance to compression of the rib cage is of interest.
Fig. D.3 shows bio-bag models of the mediastinum and heart.
The mediastinum is the complex space between the thoracic organs and vessels and
it is considered filled with an incompressible gelatinous fluid. It can be modeled by
defining a fictitious envelope, Fig. D.3(a), and by assigning the conditions of bio-bags
for quasi-incompressibility.
Similarly, the heart can be modeled by a bio-bag with a high degree of incompress-
ibility, Fig. D.3(b). For capturing the effect of expelled blood during a violent chest
impact, the modeled heart chambers can be provided with outflow vents. The thorax
models can so be calibrated to well represent the results from Kroell frontal and Viano
side pendulum impact tests, Fig. D.3(c).
Other fluid-structure interaction modeling techniques. While the described biobag

models are efficient but approximate models, the volume enclosed by a hollow organ
can be modeled with more precision with Lagrangian solid finite elements, which un-
dergo the equation of state given by a fluid, Fig. D.4, inset (a). This model is possible
when the fluid transport is small, i.e., the distortion of the Lagrange mesh is limited.
Airbag models 425
Coupled fluid-structure interaction (FSI) models (L ÖHNER [1990]), as shown for an

airbag in inset (b), will be the most accurate (and the most expensive) representations
of hollow organs. This technique is indicated if the detailed interaction of the moving
fluids with the confining wall is of interest (example: aorta rupture). In that case a fluid
code and a structure code are coupled, where the fluid code provides the wall pressure
loads to the structure code and where the structure code provides the wall positions and
velocities to the fluid code at each common solution time step. Finally, the fluid or gas
can also be represented by SPH or FPM “particle” techniques (M ONAGHAN and G IN -
GOLD [1983], M ONAGHAN [1988]), insets (c), (d). While in FSI techniques the fluid
domain is meshed, inset (b), particle methods do not require a domain mesh. This is
particularly convenient when the fluid domain connectivity changes, as in the example
of the heart valve, inset (d).
A PPENDIX E
Interactions between Parts
Contact simulation. The numerical treatment of contact with various types of contact
options was mentioned in Appendix A. Effective treatment of contact is not only of
prime importance for modeling impact biomechanics, but certain contact algorithms
serve also in assembling the complex geometries of the parts of the human body. Some
examples, taken from the HUMOS model (ESI version), explain how different modeling
strategies can provide viable solutions to the complex mechanical interactions between
organs and parts of the human body.
F IG . E.1. Connections between parts via contact (HUMOS-ESI model).
F IG . E.2. Non-matching flesh and bone meshes connected by tied contacts (HUMOS-ESI model).
427
428 E. Haug et al. A PPENDIX E
Mesh merging. Apart from simulating dynamic collisions between moving parts,
special contact options are frequently used within models of articulations, to delimit
adjacent organs and to conveniently join differently meshed parts, Fig. E.1 (B EAUGO -
NIN , C OUSIN and H AUG [2001a], B EAUGONIN , C OUSIN and H AUG [2001b]).
In another example, Fig. E.2 shows the independent meshes of the leg bones and the
surrounding flesh, B EAUGONIN and H AUG [2001].
This figure demonstrates how different constituents of body parts, such as flesh and
bone, are linked with connective membranes and tissue, and are often meshed inde-
pendently for convenience. Both material constituents are connected by a tied contact
option (type 32 in Table A.1, Appendix A).
Example: Lower extremity. To control the interaction between the different compo-
nents involved in the lower limb segment of the HUMOS model, several types of sliding
interfaces have been defined in its PAM-Crash version, Fig. E.3 (B EAUGONIN and
H AUG [2001]).
F IG . E.3. Articular, ligamant/tendon-to-bone and self-contacts in the lower extremity (HUMOS-ESI model)
(B EAUGONIN and H AUG [2001]).
Interactions between parts 429
(i) The sliding interfaces between bones near their articulations are defined with a
segment-to-segment contact (type 33 in Table A.1, Appendix A).
(ii) The sliding interfaces between bone and ligament or tendon, and between bone
and skin are defined with a segment-to-segment or a node-to-segment contact
(type 34 in Table A.1, Appendix A). If necessary, an edge-to-edge contact (type
46 in Table A.1, Appendix A) is added to avoid the penetration between the
components.
(iii) The interaction between ligaments is controlled by a segment-to-segment con-
tact or a self-contact (type 36 in Table A.1, Appendix A). The sliding interface
between ligament/tendon and skin is defined by a node-to-segment contact.
F IG . E.4. Attachments between abdominal organs (HUMOS-ESI model).

430 E. Haug et al. A PPENDIX E
Example: Abdominal organs. The abdominal organs are tied together and are loosely
tethered to the abdominal walls (peritoneum) by ligaments or folds of the peritoneum.
These features can be modeled with tied contact options in the ESI HUMOS model. The
tied contact options permit no relative motion between the tied parts, but can be allowed
to break when certain contact force limits are exceeded. This option can approximate the
rupture of ligaments and other tissue connections. Some of the connections are shown
as examples in Fig. E.4 (after B EAUGONIN , A LLAIN and H AUG [2001]).
The abdominal organs, as well as the brain and the thoracic organs, interact with
neighbors and cavity walls under considerable relative sliding motions. This can be
modeled with the standard sliding contact options.
The organs fill their host cavities with literally no voids or gaps in the sense that
the space between the organs is filled with tissue or fluids. This can be modeled ap-
proximately with “bio-bags”, as shown in Appendix D for the mediastinum, where the
F IG . E.5. Simulation of cut liver with tied meshes between vena porta and parenchyma (CAESARE-ESI
model: D AN and M ILCENT [2002]).
Interactions between parts 431
organ-to-organ contact can be modeled with slide-and-void contacts between the or-
gans, while the incompressibility of the liquid-filled space between them is assured by
the bio-bag feature (Appendix D). The same feature was applied in Chapter 3, Section 8
(Fig. 8.8) for the modeling of the Cavanaugh bar impact test on the abdomen.
Example: Liver. For practical reasons (convenience, meshing freedom, mesh size lim-
itations, etc.) vessels inside internal organs are often modeled apart from the bulk matter
of the organs. Automatic mesh merging techniques, or tied contact options, can then be
applied to tie the non-congruent meshes between the surface of the meshed vessels (of-
ten: shells) and the organ bulk matter (solids). Fig. E.5 shows a model of a human liver
with the internal arborescence of a systems of vessels (vena porta) exposed through
a simulated progressive cut into the parenchyma (from the CAESARE Project DAN
and M ILCENT [2002]). Glisson’s capsule around the liver is modeled with thin mem-
branes. The cut-in-progress was simulated with assigning almost zero resistance to the
sectioned elements. In order to expose the incision, the cut portion of the liver is not
supported by the horizontal support plate and it deflects through the action of gravity. It
is still connected by its uncut portion with the main portion of the organ. The tougher
vessels become visible by assigning transparency to the surrounding parenchyma solid
elements, which appear only through their exposed surface grids (blue color).
References
A BEL , J.M., G ENNARELLI , T.A., S EGAWA , H. (1978). Incidence and severity of cerebral concussion in the
rhesus monkey following sagittal plane angular acceleration. In: Proc. 22nd Stapp Car Crash Conference,
pp. 35–53. Paper No. 780886.
A L -B SHARAT , A.S., H ARDY , W.N., YANG , K.H., K HALIL , T.B., TASHMAN , S., K ING , A.I. (1999).
Brain-skull relative displacement magnitude due to blunt head impact: New experimental data and model.
In: Proc. 43rd Stapp Car Crash Conference, pp. 101–160. Paper No. 99SC22.
A LLAIN, J.C. (1998). Etude et calibration d’un modèle numérique de thorax, Internal Report, ESI Software
S.A., 99, rue des Solets, BP 80112, 94513 Rungis Cedex, France.
A LLEN , B.L., F ERGUSON , R.L., L EHMANN , T.R., O’B RIEN , R.P. (1982). A mechanistic classification of
closed indirect fractures and dislocations of the lower cervical spine. Spine 7 (1), 1–27.
A LLSOP , D.L. (1993). Skull and facial bone trauma: Experimental aspects. In: Nahum, A.M., Melvin, J.W.
(eds.), Accidental Injury – Biomechanics and Prevention (Springer, Berlin), pp. 247–267 (Chapter 11).
A RMSTRONG , C.G., L AI , W.M., M OW , V.C. (1984). An analysis of the unconfined compression of articular
cartilage. J. Biomech. Engrg. 106, 165–173.
A SHMAN , R.B., et al. (1984). A continuous wave technique for the measurement of the elastic properties of
cortical bone. J. Biomech., 349.
A SHMAN , R.B., et al. (1986). Ultrasonic technique for the measurement of the structural elastic modulus of
cancellous bone. Trans. Orthopedic Res. Soc., 43.
ATTARIAN , D.E., M C C RACKIN , H.J., D E V ITO , D.P., M C E LHANEY, J.H., G ARRETT, W.E. (1985). Bio-
mechanical characteristics of human ankle ligaments. Foot & Ankle 6 (2).
BACH , J.M., H ULL , M.L., PATTERSON , H.A. (1997). Direct measurement of strain in the postero-lateral
bundle of the anterior cruciate ligament. J. Biomech. 3 (3), 281–283.
BAHLER , A.S., FALES , J.T., Z IERLER , K.L. (1968). The dynamic properties of mammalian skeletal muscle.
J. Gen. Physiol. 51, 369–384.
BANDAK , F.A. (1996). Biomechanics of impact traumatic brain injury. In: Proceedings of the NATO ASI
on Crashworthiness of Transportation Systems, Structural Impact and Occupant Safety, Troia, Portugal,
pp. 213–253.
BANDAK , F.A., E PPINGER , R.H. (1994). A three-dimensional finite element analysis of the human brain
under combined translational and rotational acceleration. In: Proc. 38th Stapp Car Crash Conference,
pp. 145–163. Paper No. 942215.
BANDAK , F.A., TANNOUS , R.E., Z HANG , A.X., D I M ASI , F., M ASIELLO , P., E PPINGER , R.H. (2001).
SIMon: A simulated injury monitor, application to head injury assessment. In: 17th International Technical
Conference on the Enhanced Safety of Vehicles, Amsterdam, Holland.
BANDAK , F.A., TANNOUS , R.E., Z HANG , A.X., T ORIDIS , T.G., E PPINGER , R.H. (1996). Use of finite
element analysis and dummy test measurements for the assessment of crash injury traumatic brain injury.
In: Advisory Group for Aerospace Research and Development, Mescalero, New Mexico, pp. 10–10.13.
BARTLEY , M.H., et al. (1966). The relationship of bone strength and bone quantity in health, disease and
aging. J. Gerontol. 21, 517.
BATHE , J. (1996). Finite Element Procedures (Prentice-Hall, Englewood Cliffs, NY).
BAUDER, B. (1985). The dynamic load tolerance of the human liver (Die dynamisch mechanische Belast-
barkeit der menschlichen Leber). PhD Dissertation, Medicine, University of Heidelberg.
433
434 E. Haug et al.
BAUDRIT , P., H AMON , J., S ONG , E., ROBIN , S., L E C OZ , J.-Y. (1999). Comparative studies of dummy and
human body models in frontal and lateral impact conditions. In: Proc. 43rd Stapp Car Crash Conference,
pp. 55–75. Paper No. 99SC05.
B EAUGONIN, M., A LLAIN, J.C., H AUG, E. (2001). Pam–Crash modeling and validation: Pelvis–Abdomen
segment. HUMOS Report 5ESI/010131/T3/DA, ESI Software S.A., 99, rue des Solets, BP 80112, 94513
Rungis Cedex, France.
B EAUGONIN, M., C OUSIN, G., H AUG, E. (2001a). Pam–Crash: Whole model validation. HUMOS Report
6ESI/010131/T5/DA, ESI Software S.A., 99, rue des Solets, BP 80112, 94513 Rungis Cedex, France.
B EAUGONIN, M., C OUSIN, G., H AUG, E. (2001b). Pam–Crash modeling and validation: Thorax-shoulder
segment. HUMOS Report 6ESI/010131/T2/DA, ESI Software S.A., 99, rue des Solets, BP 80112, 94513
Rungis Cedex, France.
B EAUGONIN, M., H AUG, E. (2001). Pam–Crash modeling and validation: Lower limb segment. HUMOS
Report 5ESI/010131/T5/DA, ESI Software S.A., 99, rue des Solets, BP 80112, 94513 Rungis Cedex,
France.
B EAUGONIN , M., H AUG , E., C ESARI , D. (1996). A numerical model of the human ankle/foot under impact
loading in inversion and eversion. In: Proc. 40th Stapp Car Crash Conference Proceedings, Albuquerque,
NM, pp. 239–249. Paper No. 962428.
B EAUGONIN , M., H AUG , E., C ESARI , D. (1997). Improvements of numerical ankle/foot model: Modeling
of deformable bone. In: Proc. 41st Stapp Car Crash Conference, pp. 225–237. Paper No. 973331.
B EAUGONIN, M., H AUG, E., H YNCIK, L. (1998). Robby2 – alpha-version. Internal Report, ESI Software
S.A., 99, rue des Solets, BP 80112, 94513 Rungis Cedex, France.
B EAUGONIN , M., H AUG , E., M UNCK , G., C ESARI , D.A. (1995). Preliminary numerical model of the hu-
man ankle under impact loading. In: Pelvic and Lower Extremity Injuries (PLEI) Conference, Washington,
DC, USA.
B EAUGONIN , M., H AUG , E., M UNCK , G., C ESARI , D. (1996). The influence of some critical parameters on
the simulation of the dynamic human ankle dorsiflexion response. In: 15th ESV Conference, 96-S10-W-31,
Melbourne.
B EDEWI , P.G., M IYAMOTO, N., D IGGES , K.H., B EDEWI , N.E. (1998). Human femur impact and injury
analysis utilizing finite element modeling and real-world case study data. In: PUCA ’98 Proceedings,
Nihon ESI, ESI Software S.A., 99, rue des Solets, BP 80112, 94513 Rungis Cedex, France, vol. 1, pp. 311–
320.
B EGEMAN , P., A EKBOTE , K. (1996). Axial load strength and some ligament properties of the ankle joint. In:
Injury Prevention Through Biomechanics Symposium Proceedings, Hutzel Hospital, Wayne State Univer-
sity, Detroit, Michigan, USA, pp. 125–135.
B EGEMAN , P., BALAKRISHNAN , P., L EVINE , R., K ING , A. (1992). Human ankle response in dorsiflexion.
In: Injury Prevention Through Biomechanics Symposium Proceedings, Wayne State University.
B EGEMAN , P., BALAKRISHNAN , P., L EVINE , R., K ING , A. (1993). Dynamic human ankle response to inver-
sion and eversion. In: Proc. 37th Stapp Car Crash Conference Proceedings, pp. 83–93. Paper No. 933115.
B EGEMAN , P.C., K ING , A.I., P RASAD , P. (1973). Spinal loads resulting from −gx acceleration. In: Proc.
17th Stapp Car Crash Conference, Warrendale, PA. Paper No. 730977.
B EGEMAN , P., K OPACZ , J.M. (1991). Biomechanics of human ankle impact in dorsiflexion. In: Injury Pre-
vention Through Biomechanics Symposium Proceedings, Wayne State University.
B EHRENS , J.C., et al. (1974). Variation in strength and structure of cancellous bone at the knee. J. Biomech. 7,
201–207.
B EILLAS , P., B EGEMAN , P.C., YANG , K.H., K ING , A., A RNOUX , P.-J., K ANG , H.-S., K AYVANTASH , K.,
B RUNET , C., C AVALLERO , C., P RASAD , P. (2001). Lower limb: Advanced FE model and new experi-
mental data. Stapp Car Crash J. 45, 469–494. Paper No. 2001-22-0022.
B EILLAS , P., L AVASTE , F., N ICOLOPOULOS , D., K AYVANTASH , K., YANG , K.H., ROBIN , S. (1999). Foot
and ankle finite element modeling using CT-scan data. In: Proc. 43rd Stapp Car Crash Conference, pp. 1–
14. Paper No. 99SC11.
B ELYTSCHKO , T., L IN , J.I. (1984). Explicit algorithms for the nonlinear dynamics of shells. Comput. Meth-
ods Appl. Mech. Engrg. 42, 225–251.
B ELYTSCHKO , T., K ULAK , R.F., S CHULTZ , A.B., G ALANTE , J.O. (1972). Numerical stress analysis of
intervertebral disk. In: ASME (Biomechanical & Human Factors Division), Winter Annual Meeting, NY.
References 435
B ELYTSCHKO , T., K ULAK , R.F., S CHULTZ , A.B., G ALANTE , J.O. (1974). Finite element stress analysis of
an intervertebral disc. J. Biomech. 7, 277–285.
B ELYTSCHKO , T., T SAY , C.S. (1983). A stabilization procedure for the quadrilateral plate element with
one-point quadrature. Internat. J. Numer. Methods Engrg. 19, 405–419.
B ELYTSCHKO , T., W ONG , J.S., L IU , W.K., K ENNEDY , J.M. (1984). Hourglass control in linear and non-
linear problems. Comput. Methods Appl. Mech. Engrg. 43, 251–276.
B ENES , K. (2002). ROBBY 2 Model with Muscles, Internal Report, ESI Software S.A., 99, rue des Solets,
BP 80112, 94513 Rungis Cedex, France.
B LACK , J., H ASTINGS , G. (eds.) (1998). Handbook of Biomaterial Properties, Part I (Chapman & Hall,
London).
B OKDUK , N., Y OGANANDAN , N. (2001). Biomechanics of the cervical spine, Part 3: Minor injuries. Clinical
Biomech. 16, 267–275.
B OUISSET , S., M ATON , B. (1995). Muscles, Posture et Mouvement (Hermann, Paris).
B OUQUET , R., R AMET , M., B ERMOND , F., C ESARI , D. (1994). Thoracic and human pelvis response to
impact. In: Proceedings of the 14th International Technical Conference on Enhanced Safety of Vehicles
(ESV), pp. 100–109.
B OWMAN , B.M., S CHNEIDER , L.W., L USTAK , L.S., A NDERSON , W.R., T HOMAS , D.J. (1984). Simula-
tion analysis of head and neck dynamic response. In: Proc. 28th Stapp Car Crash Conference, pp. 173–205.
Paper No. 841668.
B ROWN , T.D., et al. (1980). Mechanical property distribution in the cancellous bone of the human proximal
femur. Acta Orthop. Scand. 51, 429–437.
B SCHLEIPFER, Th. (2002). Das experimentelle stumpfe Nierentrauma: Biomechanik, Traumaverhalten und
bildgebende Diagnostik. M.D. Dissertation, Faculty of Medicine, University Ulm, Germany.
B URGHELE, N., S CHULLER, K. (1968). Die Festigkeit der Knochen Kalkaneus und Astragallus. Aus der
Klinik fur Orthopädie des Bukarester Unfallkrankenhauses und dem Lehrstuhl fur Festigkeit und Mate-
rialprufungswesen der Technischen Hochschule, Bukarest, Rumänien (in German).
B URSTEIN , A.H., et al. (1976). Aging of bone tissue: Mechanical properties. J. Bone Joint Surgery A 58, 82.
C ARTER , D.R., H AYES , W.C. (1977). The compressive behavior of bone as a two-phase porous structure.
J. Bone Joint Surgery A 59, 954–962.
C ARTER, F. (1999). Work at the University of Dundee. Private communication to INRIA, Sophia-Antipolis.
C AVANAUGH , J.M. (1993). The biomechanics of thoracic trauma. In: Nahum, A.M., Melvin, J.W. (eds.),
Accidental Injury – Biomechanics and Prevention (Springer, Berlin), pp. 362–390 (Chapter 15).
C AVANAUGH , J.M., N YQUIST , G.W., G OLDBERG , S.J., K ING , A.I. (1986). Lower abdominal tolerance
and response. In: Proc. 30th Stapp Car Crash Conference, pp. 41–63. Paper No. 861878.
C AVANAUGH , J.M., WALILKO , T.J., M ALHOTRA , A., Z HU , Y., K ING , A.I. (1990). Biomechanical re-
sponse and injury tolerance of the pelvis in twelve sled side impact tests. In: Proc. 34th Stapp Car Crash
Conference, pp. 23–38. Paper No. 902307.
C ESARI , D., B ERMOND, F., B OUQUET, R., R AMET, M. (1994). Virtual predictive testing of biomechanical
effects of impacts on the human leg. In: Proceedings of PAM ’94, ESI Software S.A., 99, rue des Solets,
BP 80112, 94513 Rungis Cedex, France, pp. 177–188.
C ESARI , D., B OUQUET , R. (1990). Behavior of human surrogates thorax under belt loading. In: Proc. 34th
Stapp Car Crash Conference, pp. 73–81. Paper No. 902310.
C ESARI , D., B OUQUET , R. (1994). Comparison of Hybrid III and human cadaver thorax deformation loaded
by a thoracic belt. In: Proc. 38th Stapp Car Crash Conference, pp. 65–76. Paper No. 942209.
C HANCE , G.O. (1948). Note on a type of flexion fracture of the spine. Br. J. Radiol. 21, 452–453.
C HANDLER , R.F. (1993). Development of crash injury protection in civil aviation. In: Nahum, A.M., Melvin,
J.W. (eds.), Accidental Injury – Biomechanics and Prevention (Springer, Berlin), pp. 151–185 (Chapter 6).
C HAPON, A., V ERRIEST, J.P., D EDOYAN, J., T RAUCHESSEC, R., A RTRU, R. (1983). Research on brain
vulnerability from real accidents, ISO Document No. ISO/TC22SC12/GT6/N139.
C HAZAL , J., TANGUY, A., B OURGES , M. (1985). Biomechanical properties of spinal ligaments and a his-
torical study of the supraspinal ligament in traction. J. Biomech. 18 (3), 167–176.
C HOI , H.-Y. (2001). Numerical human head model for traumatic injury assessment. KSME Internat. J. 15 (7),
995–1001.
436 E. Haug et al.
C HOI , H.-Y., E OM , H.-W. (1998). Finite element modeling of human cervical spine. Hongik J. Sci. Technol.
(Special volume).
C HOI , H.-Y., L EE, I.-H. (1999a). Finite element modeling of human thorax for occupant safety simulation.
In: 10th Proceedings of PUCA, Japan; ESI Software S.A., 99, rue des Solets, BP 80112, 94513 Rungis
Cedex, France, pp. 469–478.
C HOI , H.-Y., L EE, I.-H. (1999b). Advanced finite element modeling of the human body for occupant safety
simulation. In: EuroPam 1999, ESI Software S.A., 99, rue des Solets, BP 80112, 94513 Rungis Cedex,
France.
C HOI , H.-Y., L EE , I.-H., E OM , H.-W., L EE , T.-H. (1999). A study on the whiplash injury due to the low
velocity rear-end collision. In: 3rd World Congress of Biomechanics, Japan, p. 522.
C HOI , H.-Y., L EE , I.-H, H AUG , E. (2001a). Finite element modeling of human upper extremity for occupant
safety simulation. In: JSAE Spring Conference, No. 34-01, 20015358.
C HOI , H.-Y., L EE , I.-H., H AUG , E. (2001b). Finite element modeling of human head–neck complex for
crashworthiness simulation. In: First MIT Conference Proceedings.
C IARELLI , M.J., et al. (1986). Experimental determination of the orthogonal mechanical properties, density,
and distribution of human trabecular bone from the major metaphyseal regions utilizing material testing
and computed tomography. Trans. Orthop. Res. Soc., 42.
C IARELLI , M.J., et al. (1991). Evaluation of orthogonal mechanical properties and density of human trabec-
ular bone from the major metaphyseal regions with materials testing and computed tomography. J. Orth.
Res. 9, 674–682.
C LAESSENS , M. (1997). Finite Element Modelling of the Human Head under Impact Conditions. PhD Thesis.
Eindhoven University. ISBN 90-386-0369-X.
C LAESSENS , M., S AUREN , F., W ISMANS , J. (1997). Modeling the human head under impact conditions:
A parametric study. In: Proc. 41st Stapp Car Crash Conference, pp. 315–328. Paper No. 973338.
C LEMENTE , C.D. (1981). Anatomy (A Regional Atlas of the Human Body), second ed. (Urban & Schwarzen-
berg, Baltimore–Munich).
C OHEN , D.S. (1987). The safety problem for passengers in frontal impacts: Analysis of accidents, labora-
tory and model simulation data. In: 11th ESV International Technical Conference on Experimental Safety
Vehicles, Washington, DC.
C OLE , G.K., VAN DEN B OGERT , A.J., H ERZOG , W., G ERRITSEN , K.G.M. (1996). Modelling of force
prediction in skeletal muscle undergoing stretch. J. Biomech. 29 (8), 1091–1104.
C OMPTON , C.P. (1993). The use of public crash data in biomechanical research. In: Nahum, A.M., Melvin,
C OOPER , P.R. (1982a). Post-traumatic intracranial mass lesions. In: Cooper, P.R. (ed.), Head Injury
(Williams and Wilkins, Baltimore/London), pp. 185–232.
C OOPER , P.R. (1982b). Skull fracture and traumatic cerebrospinal fluid fistulas. In: Cooper, P.R. (ed.), Head
Injury (Williams and Wilkins, Baltimore/London), pp. 65–82.
C OOPER , G.J., P EARCE , B.P., S TAINER , M.C., M AYNARD , R.L. (1982). The biomechanical response of
the thorax to non-penetrating trauma with particular reference to cardiac injuries. J. Trauma 22 (12), 994–
1008.
C OSTA -PAZ, M., R ANALLETTA, M., M AKINO, A., AYERZA, M., M USCOLO, L. (2002). Displaced patella
fracture after cruciate ligament reconstruction with patellar ligament graft, SICOT Online Report E006
February 2002, http://www.sicot.org/.
C RANDALL , J.R., D UMA , S.M., BASS , C.R., P ILKEY , W.D., K UPPA , S.M., K HAEWPONG , N., E P -
PINGER , R. (1999). Thoracic response and trauma in airbag deployment tests with out-of-position small
female surrogates. J. Crash Prevention Injury Control 1 (2), 101–102.
C RAWFORD , G.N.C., JAMES , N.T. (1980). The design of muscles. In: Owen, R., Goodfellow, J., Bullough, P.
(eds.), Scientific Foundations of Orthopaedics and Traumatology (William Heinemann, London), pp. 67–
74.
C URREY , J.D. (1975). The effects of strain rate, reconstruction and mineral content on some mechanical
properties of bovine bone. J. Biomech., 81.
C USICK , J.F., Y OGANANDAN , N. (2002). Biomechanics of the cervical spine 4: Major injuries. Clinical
Biomech. 17, 1–20.
D ALSTRA , M., H UISKES , R. (1995). Load transfer across the pelvic bone. J. Biomech. 6 (2), 715–724.
References 437
D ALSTRA , M., et al. (1993). Mechanical and textural properties of pelvic trabecular bone. J. Biomech. 26 (4–
5), 523–535.
D AN, D. (1995). Elaboration d’un capteur de pression. Application au foie humain. Projet de fin d’étude
D.E.S.S., Collaboration entre L’Université Paris 7 et le LAB Laboratoire d’Accidentologie, de Biomé-
canique et d’Etude du Comportement Humain de PSA Peugeot Citroën–Renault.
D AN, D. (1999). Caractérisation mécanique du foie humain en situation de choc. PhD Thesis, University Paris
7 – Denis Diderot.
D AN, D., M ILCENT, G. (2002). Caesare – chirurgie abdominale et simulation a retour d’effort, French Min-
istry of Education, Research and Technology Project, Final Technical Report RE/02.1600/A, Contract
FSP9E2045, ESI France, 99 rue des Solets, 94513 Rungis Cedex, France.
D AVIDSSON , J., F LOGARD , A., L ÖVSUND , P., S VENSSON , M.Y. (1999). BioRID P3—Design and perfor-
mance compared to Hybrid III and Volunteers in Rear Impacts of V = 7 km/h. In: Proc. 43rd Stapp Car
Crash Conference, pp. 253–265. Paper No. 99SC16.
D ENG, Y.C. (1985). Human head/neck/upper-torso model response to dynamic loading, PhD Thesis, Univer-
sity of California.
D ENG , Y.C., G OLDSMITH , W. (1987). Response of a human head/neck/upper-torso replica to dynamic load-
ing – II. Analytical/numerical model. J. Biomech. 20, 487–497.
D IGIMATION /V IEWPOINT C ATALOG (2002). External and Skeletal Anatomy, 2002 Edition.
D I M ASI , F., M ARCUS , J., E PPINGER , R. (1991). 3-D anatomic brain model for relating cortical strains
to automobile crash loading. In: 13th ESV International Technical Conference on Experimental Safety
Vehicles. Paper No. 91-S8-O-11.
D OHERTY , B.J., E SSES , S.L., H EGGENESS , M.H. (1992). A biomechanical study of odontoid fractures and
fracture fixation. In: Cervical Spine Research Society.
D ONNELLY , B.R., M EDIGE , J. (1997). Shear properties of human brain tissue. J. Biomech. Engrg. 119,
423–432.
D OSTAL , W.F. (1981). A three-dimensional biomechanical model of hip musculature. J. Biomech. 14 (11),
803–812.
D UCHEYNE , P., et al. (1977). The mechanical behavior of intracondylar cancellous bone of the femur at
different loading rates. J. Biomech. 10, 747–762.
D VORAK , J., H AYEK , J., Z EHNUDER , R. (1987). CT-functional diagnostics of rotary instability of upper cer-
vical spine. Part II: An evaluation on healthy adults and patients with suspected instability. J. Spine 12 (8),
726.
D VORAK , J., PANJABI , M.M., F ROEHLICH , D., et al. (1988). Functional radiographic diagnosis of the cer-
vical spine: Flexion/extension. J. Spine 13 (7), 748.
E BARA , S., I ATRIDIS , J.C., S ETTON , L.A., F OSTER , R.J., M OW , V.C., W EIDENBAUM , M. (1996). Tensile
properties of nondegenerate human lumbar anulus fibrosus. J. Spine 21 (4), 452–461.
E NGIN , A.E. (1979). Passive resistance torques about long bone axes of major human joints. Aviation, Space
and Environmental Medicine 50 (10), 1052–1057.
E NGIN , A.E. (1980). On the biomechanics of the shoulder complex. J. Biomech. 13-7, 575–590.
E NGIN , A.E. (1983). Dynamic modelling of human articulating joints. Math. Modelling 4, 117–141.
E NGIN , A.E. (1984). On the damping properties of the shoulder complex. J. Biomech. Engrg. 106, 360–363.
E NGIN , A.E., C HEN , S.M. (1988a). On the biomechanics of the human hip complex in vivo – I. Kinematics
for determination of the maximal voluntary hip complex sinus. J. Biomech. 21 (10), 785–796.
E NGIN , A.E., C HEN , S.M. (1988b). On the biomechanics of the human hip complex in vivo – II. Passive
resistive properties beyond the hip complex sinus. J. Biomech. 21 (10), 797–806.
E NGIN , A.E., C HEN , S.M. (1989). A statistical investigation of the in vivo biomechanical properties of the
human shoulder complex. Math. Comput. Modelling 12 (12), 1569–1582.
E NGIN , A.E., P EINDL , R.D. (1987). On the biomechanics of human shoulder complex – I. Kinematics for
determination of the shoulder complex sinus. J. Biomech. 20 (2), 103–117.
E NGIN , A.E., T’ MER , S.T. (1989). Three-dimensional kinematic modelling of the human shoulder complex
– Part I: Physical model and determination of joint sinus cones. J. Biomech. Engrg. 111, 107–112.
E PPINGER , R.H. (1976). Prediction of thoracic injury using measurable experimental parameters. In: 6th
ESV International Technical Conference on Experimental Safety Vehicles, pp. 770–779.
438 E. Haug et al.
E PPINGER , R.H. (1993). Occupant restraint systems. In: Nahum, A.M., Melvin, J.W. (eds.), Accidental Injury
– Biomechanics and Prevention (Springer, Berlin), pp. 186–197 (Chapter 8).
E PPINGER , R.H., M ARCUS , J.H., M ORGAN , R.M. (1984). Development of dummy and injury index for
NHTSA’s thoracic side impact protection research program. In: Government/Industry Meeting and Expo-
sition, Washington, DC. SAE 840885.
E STES , M.S., M C E LHANEY, J.H. (1970). Response of brain tissue to compressive loading. ASME Paper
No. 70-BHF-13.
E VANS , F.G., et al. (1961). Regional differences in some physical properties of human spongy bone. In:
Evans, F.G. (ed.), Biomechanical Studies of the Musculo–Skeletal System (CC Thomas, Springfield, IL),
pp. 49–67.
E WING , C., T HOMAS , D., L USTICK , L., M UZZY , W. III, W ILLEMS , G., M AJEWSKI , P. (1976). The effect
of duration, rate of onset, and peak sled acceleration on the dynamic response of the human head and neck.
In: Proc. 20th Stapp Car Crash Conference, pp. 3–41. Paper No. 760800.
E WING , C., T HOMAS , D., L USTICK , L., M UZZY , W. III, W ILLEMS , G., M AJEWSKI , P. (1978). Effect of
initial position on the human head and neck response to +Y impact acceleration. In: Proc. 22nd Stapp Car
Crash Conference, pp. 103–138. Paper No. 780888.
FALLENSTEIN , G.T., H ULCE , V.D., M ELVIN , J.W. (1970). Dynamic mechanical properties of human brain
tissue. J. Biomech. 2, 217–226.
FAZEKAS , I.G., K OSA , F., J OBBA , G., M ESZARO , E. (1971a). Die Druckfestigkeit der Menschlichen Leber
mit besonderer Hinsicht auf Verkehrsunfälle. Z. Rechtsmedizin 68, 207–224.
FAZEKAS , I.G., K OSA , F., J OBBA , G., M ESZARO , E. (1971b). Experimentelle Untersuchungen über die
Druckfestigkeit der Menschlichen Niere. Zacchia 46, 294–301.
FAZEKAS , I.G., K OSA , F., J OBBA , G., M ESZARO , E. (1972). Beiträge zur Druckfestigkeit dser Men-
schlichen Milz bei stumpfen Gewalteinwirkungen. Arch. Kriminol. 149, 158–174.
F IROOZBAKSHK (1975). A model of brain shear under impulsive torsional loads. J. Biomech. 8, 65–73.
F OGRASCHER, K. (1998). Development of a simulation model for the head protection system ITS and inte-
gration of the component model into the full structural vehicle model, EuroPam 1998, ESI Software S.A.,
99, rue des Solets, BP 80112, 94513 Rungis Cedex, France.
F OREMAN , S.M., C ROFT , A.C. (1995). Whiplash Injuries: The cervical Acceleration/Deceleration Syn-
drome, second ed. (Williams & Wilkins).
F RANK , C.B., S HRIVE , N.G. (1994). Biomechanics of the musculo–skeletal System – 2.3: Ligament. In:
Nigg, B.M., Herzog, W. (eds.), Biomechanics of the Musculo–Skeletal System (University of Calgary,
Alberta, Canada), pp. 106–132.
F RANKEL , V.H., N ORDIN , M. (1980). Basic Biomechanics of the Skeletal System (Lea & Febiger, Philadel-
phia, PA).
F UKUBAYASHI , T., K UROSAWA , H. (1980). The contact area and pressure distribution pattern of the knee.
J. Acta Orthop. Scand. 51, 871–879.
F UNG , Y.C. (1993a). Biomechanics – Mechanical Properties of Living Tissues (Springer, Berlin).
F UNG , Y.C. (1993b). The application of biomechanics to the understanding of injury and healing. In: Nahum,
A.M., Melvin, J.W. (eds.), Accidental Injury – Biomechanics and Prevention (Springer, Berlin), pp. 1–11
(Chapter 1).
F UNG, Y.C., Y EN, M.R. (1984). Experimental investigation of lung injury mechanisms, Topical Report, US
Army Medical Research and Development Command. Contract No. DAMD 17-82-C-2062.
F URUKAWA , K., F URUSU , K., M IKI , K. (2002). A development of child FEM model – Part I: Skeletal model
of six-year-old child. In: Proc. 2002 JSME Annual Congress, No. 02-1, pp. 89–90 (in Japanese).
G ADD , C.W. (1961). Criteria for injury potential. In: Impact Acceleration Stress Symposium, National Re-
search Council Publication, No. 977 (National Academy of Sciences, Washington, DC), pp. 141–144.
G ADD , C.W. (1966). Use of a weighted impulse criterion for estimating injury hazard. In: Proc. 10th Stapp
Car Crash Conference, pp. 164–174. Paper No. 660793.
G ALANTE , J.O. (1967). Tensile properties of human lumbar anulus fibrosus. Acta Orthop Scand (Suppl.) 100.
G ALANTE , J.O., et al. (1970). Physical properties of trabecular bone. Calcif. Tissue Res. 5, 236–246.
G ALFORD , J.E., M C E LHANEY , J.H. (1970). A visco-elastic study of scalp, brain and dura. J. Biomech. 3,
211–221.
References 439
G ENNARELLI , T.A. (1980). Analysis of head injury severity by AIS-80. In: 24th Annual Conference of the
American Association of Automotive Medicine (AAAM, Morton Grove, IL), pp. 147–155.
G ENNARELLI , T.A., T HIBAULT , L.E. (1982). Biomechanics of acute subdural hematoma. J. Trauma 22 (8),
680–686.
G ENNARELLI , T.A., T HIBAULT , L.E., A DAMS , J.H., G RAHAM , D.I., T HOMPSON , C.J., M ARCINCIN ,
R.P. (1982). Diffuse axonal injury and traumatic coma in the primate. Ann. Neuron. 12, 564–574.
G ENNARELLI , T.A., T HIBAULT , L.E., T OMEI , G., W ISER , R., G RAHAM , D., A DAMS , J. (1987). Direc-
tional dependence of axonal brain injury due to centroidal and non-centroidal acceleration. In: Proc. 31st
G OEL , V.K., G OYAL , S., C LARK , C., N ISHIYAMA , K., N YE , T. (1985). Kinematics of the whole lumbar
spine effect of dissectomy. J. Spine 10 (6).
G OEL , V.K., M ONROE , B.T., G ILBERTSON , L.G., B RINKMAN , P. (1995). Interlaminar shear stresses and
laminae separation in a disc: finite element analysis of the L3–L4 motion segment subjected to axial
compressive loads. Spine 20 (6), 689–698.
G OLDSMITH (1972). Biomechanics of Head Injury. In: Fung, Y.C., Perrone, N., Anliker, M. (eds.), Biome-
chanics: It’s Foundation and Objectives (Prentice-Hall, Englewood Cliffs, NJ).
G OLDSTEIN , S., F RANKENBURG , E., K UHN , J. (1993). Biomechanics of bone. In: Nahum, A.M., Melvin,
G OLDSTEIN , S.A., et al. (1983). The mechanical properties of human tibia trabecular bone as a function of
metaphyseal location. J. Biomech. 16, 965–969.
G RANIK , G., S TEIN , I. (1973). Human ribs: static testing as a promising medical application. J. Biomech. 6,
237–240.
G RAY ’ S A NATOMY (1989). Williams, P.L., Warwick, R., Dyson, M., Bannister, L. (eds.), Thirty-seventh ed.,
Churchill Livington, ISBN 0 443 02588 6, after Figure 5.8A, page 554.
G URDJIAN , E.S., L ISSNER , H.R. (1944). Mechanism of head injury as studied by the cathode ray oscillo-
scope, Preliminary report. J. Neurosurgery 1, 393–399.
G URDJIAN , E.S., ROBERTS , V.L., T HOMAS , L.M. (1966). Tolerance curves of acceleration and intracranial
pressure and protective index in experimental head injury. J. Trauma 6, 600–604.
G URDJIAN , E.S., W EBSTER , J.E., L ISSNER , H.R. (1955). Observations of the mechanism of brain concus-
sion, contusion, and laceration. Surgery Gynecol. Obstet. 101, 680–690.
H AJJI , M.A., W ILSON , T.A., L AI -F OOK , S.J. (1979). Improved measurements of shear modulus and pleural
membrane tension of the lung. J. Appl. Physiol. 47, 175–181.
H ALLQUIST , J.O., G OUDREAU , G.L., B ENSON , D.J. (1985). Sliding surfaces with contact–impact in large-
scale Lagrangian computations problems. Comput. Methods Appl. Mech. Engrg. 51, 107–137.
H ARDY , W.N. (1993). Instrumentation in experimental design. In: Nahum, A.M., Melvin, J.W. (eds.), Acci-
dental Injury – Biomechanics and Prevention (Springer, Berlin), pp. 12–48 (Chapter 2).
H ARDY , W.N., F OSTER , C.D., M ASON , M.J., YANG , K.H., K ING , A.I., TASHMAN , S. (2001). Investi-
gation of head injury mechanisms using neutral density technology and high-speed biplanar X-ray. Stapp
Car Crash Conference J. 45, 337–368. Paper No. 2001-22-0016.
H ARRY , J.D., WARD , A.W., H EGLUD , N.C., M ORGAN , D.L., M C M AHON , T.A. (1990). Cross-bridge
cycling theories cannot explain high-speed lengthening behavior in frog muscle. Biophys. J. 57, 201–208.
H AUG, E. (1995). Biomechanical models in vehicle accident simulation. In: PAM – User’s Conference in
Asia, PUCA ’95, Shin-Yokohama, ESI Software S.A., 99, rue des Solets, BP 80112, 94513 Rungis Cedex,
France, pp. 233–256.
H AUG, E., B EAUGONIN, M., T RAMECON, A., H YNCIK, L. (1999). Current status of articulated and de-
formable human models for impact and occupant safety simulation at ESI group. In: European Conference
on Computational Mechanics, EEVC’99, August 31 – September 3, 1999, Munic.
H AUG , E., C LINCKEMAILLIE , J.C., A BERLENC , F. (1989a). Computational mechanics in crashworthiness
analysis. In: Post Symposium Short Course of the 2nd Symposium of Plasticity, Nagoya, Japan.
H AUG , E., C LINCKEMAILLIE , J.C., A BERLENC , F. (1989b). Contact–impact problems for crash. In: Post
Symposium Short Course of the 2nd Symposium of Plasticity, Nagoya, Japan.
H AUG, E., L ASRY, D., G ROENENBOOM, P., M UNCK, G., ROGER, J., S CHLOSSER, J., RÜCKERT, J. (1993).
Finite element models of dummies and biomechanical applications using PAM-CRASH™. In: PAM –
440 E. Haug et al.
User’s Conference in Asia, PUCA ’93, Shin-Yokohama, ESI Software S.A., 99, rue des Solets, BP 80112,
94513 Rungis Cedex, France, pp. 123–158.
H AUG , E., T RAMECON , A., A LLAIN , J.C., C HOI , H.-Y. (2001). Modeling of ergonomics and muscular
comfort. KSME Internat. J. 15 (7), 982–994.
H AUG , E., U LRICH , D. (1989). The PAM–CRASH code as an efficient tool for crashworthiness simulation
and design. In: Second European Cars/Trucks Simulation Symposium, Schliersee (Munich), AZIMUTH
(Springer, Berlin).
H AUT , R.C. (1993). Biomechanics of soft tissues. In: Nahum, A.M., Melvin, J.W. (eds.), Accidental Injury –
Biomechanics and Prevention (Springer, Berlin), pp. 224–246 (Chapter 10).
H AWKINS , D., B EY , M. (1994). A comprehensive approach for studying muscle-tendon mechanics. J. Bio-
mech. Engrg. 116, 51.
H AWKINS , D., B EY , M. (1997). Muscle and tendon force-length properties and their interactions in vivo.
J. Biomech. 30 (1), 63–70.
H AYASHI , S., C HOI , H.-Y., L EVINE , R.S., YANG , K.H., K ING , A.I. (1996). Experimental and analyti-
cal study of a frontal knee impact. In: Proc. 40th Stapp Car Crash Conference, pp. 161–173. Paper No.
962423.
H ERZOG , W. (1994). Muscle. In: Nigg, B.M., Herzog, W. (eds.), Biomechanics of the Musculo–Skeletal
System (Wiley, New York), pp. 154–187.
H ILL , A.V. (1970). First and Last Experiments in Muscle Mechanics (Cambridge).
H IRSCH , C. (1955). The reaction of intervertebral discs to compression force. J. Bone Joint Surgery A 37 (6).
H OLBOURN , A.H.S. (1943). Mechanics of head injury. Lancet 2, 438–441.
H OPPIN , F.G., L EE , G.C., D AWSON , S.V. (1975). Properties for lung parenchyma in distortion. J. Appl.
Physiol. 39, 742–751.
H ORST , M.J. VAN DER, T HUNNISSEN , J.G.M., H APPEE , R., H AASTER VAN , R.M.H.P., W ISMANS ,
J.S.H.M. (1997). The influence of muscle activity on head–neck response during impact. In: Proc. 41st
H UANG , Y., K ING , A.I., C AVANAUGH , J.M. (1994a). A MADYMO model of near-side human occupants in
side impacts. J. Biomech. Engrg. 116, 228–235.
H UANG , Y., K ING , A.I., C AVANAUGH , J.M. (1994b). Finite element modelling of gross motion of human
cadavers in side impact. In: Proc. 38th Stapp Car Crash Conference, pp. 35–53. Paper No. 942207.
H UELKE , D.F., N USHOLTZ , G.S., K AIKER , P.S. (1986). Use of quadruped models in thoraco-abdominal
biomechanics research. J. Biomech. 19 (12), 969–977.
H UGHES , T.J.R., P ISTER , J.S., TAYLOR , R.L. (1979). Implicit–explicit finite elements in non-linear tran-
sient analysis. Comput. Methods Appl. Mech. Engrg. 17/18, 159–182.
H UGHES , T.J.R., TAYLOR , R.L., S ACKMAN , J.L., C URNIER , A., K ANOKNUKULCHAI , W. (1976). A finite
element method for a class of contact–impact problems. Comput. Methods Appl. Mech. Engrg. 8, 249–276.
H VID , I., et al. (1985). Trabecular bone strength patterns at the proximal tibial epiphysis. J. Orthop. Res. 3,
464–472.
H YNCIK, L. (1997). Human articulated rigid body model (ROBBY1). ESI Group Internal Report, ESI Soft-
ware S.A., 99, rue des Solets, BP 80112, 94513 Rungis Cedex, France.
H YNCIK , L. (1999a). Human shoulder model. In: 15th Conference Computational Mechanics, pp. 117–124.
ISBN 80-7082-542-1 (in Czech).
H YNCIK , L. (1999b). Human inner organs model. In: 15th Conference Computational Mechanics, pp. 125–
132. ISBN 80-7082-542-1.
H YNCIK , L. (2000). Biomechanical human model. In: 8th Conference Biomechanics of Man, pp. 52–55.
ISBN 80-244-0193-2.
H YNCIK , L. (2001a). Rigid body based human model for crash test purposes. Engrg. Mech. 5, 337–342.
H YNCIK , L. (2001b). Biomechanical model of human inner organs and tissues. In: 17th Conference Compu-
tational Mechanics, pp. 113–120. ISBN 80-7082-780-7.
H YNCIK , L. (2002a). Multi-body human model. In: Preprint of seminar Virtual Nonlinear Multibody Systems.
In: Preprint of NATO Advanced Study Institute 1, pp. 89–94.
H YNCIK , L. (2002b). Deformable human abdomen model for crash test purposes. In: 4th Conference Applied
Mechanics, pp. 159–164. ISBN 80-248-0079-9.
References 441
H YNCIK, L. (2002c). Biomechanical model of abdominal inner organs and tissues for crash test purposes. PhD
Thesis, Department of Mechanics of the Faculty of Applied Sciences of the University of West Bohemia
in Pilsen.
I RWIN, A.L. (1994), Analysis and CAL3D model of the shoulder and thorax response of seven cadavers
subjected to lateral impacts. PhD Thesis, Wayne State University, 1994.
I RWIN , A., M ERTZ , H.J. (1997). Biomechanical basis for the CRABI and Hybrid III child dummies. In:
Proc. 41st Stapp Car Crash Conference, pp. 261–272. Paper No. 973317.
I TOI , E., B ERGLUND , L.J., G RABOWSKI , J.J., S CHULTZ , F.M., G ROWNEY , E.S., M ORREY , B.F., A N ,
K.N. (1995). Tensile properties of the supraspinatus tendon. J. Orthop. Res. 13 (4), 578–584.
I WAMOTO , M., K ISANUKI , Y., WATANABE , I., F URUSU , K., M IKI , K., H ASEGAWA , J. (2002). Develop-
ment of a finite element model of the total human model for safety (THUMS) and application to injury
reconstruction. In: 2002 International IRCOBI Conference, pp. 31–42.
JAGER, M. DE (1996). Mathematical head–neck models for acceleration impact. Ph.D. Thesis, Eindhoven
University.
JAGER , M. DE, S AUREN , A., T HUNNISSEN , J., W ISMANS , J. (1994). A three-dimensional neck model: Val-
idation for frontal and lateral impact. In: 38th Stapp Car Crash Conference, pp. 93–109. Paper No. 942211.
J OHNSON , G.A., T RAMAGLINI , D.M., L EVINE , R.E., O NO , K., C HOI , N.Y., W OO , S.L. (1994). Tensile
and viscoelastic properties of human patellar tendon. J. Orthop. Res. 12 (6), 796–803.
K AJZER, J. (1991). Impact biomechanics of knee injuries. Doctoral Thesis, Dept. of Injury Prevention,
Chalmers University of Technology.
K AJZER , J., S CHROEDER , G., I SHIKAWA , H., M ATSUI , Y., B OSCH , U. (1997). Shearing and bending effects
at the knee joint at high speed lateral loading. In: Proc. 41st Stapp Car Crash Conference, pp. 151–165.
Paper No. 973326.
K AJZER, J., Z HOU, C., K HALIL, T.B., K ING, A.I. (1996). Modelling of ligaments and muscles under transient
loads: application of PAM–CRASH material models. In: PAM User’s Conference in Asia PUCA ’96, Nihon
ESI, ESI Software S.A., 99, rue des Solets BP 80112, 94513 Rungis Cedex, France, pp. 223–231.
K ALLIERIS , D., S CHMIDT , G. (1990). Neck response and injury assessment using cadavers and the US-SID
for far-side lateral impacts of rear seat occupants with inboard-anchored shoulder belts. In: Proc. 34th
K ANG , H.S., W ILLINGER , R., D IAW , B.M., C HINN , B. (1997). Validation of a 3D anatomic human head
model and replication of head impact in motorcycle accident by finite element modeling. In: Proc. 41st
K APANDJI , I.A. (1974a). The Physiology of Joints, vol. 1: Upper Limb (Churchill Livingstone).
K APANDJI , I.A. (1974b). The Physiology of Joints, vol. 2: Lower Limb (Churchill Livingstone).
K APANDJI , I.A. (1974c). The Physiology of Joints, vol. 3: The Trunk and the Vertebral Column (Churchill
Livingstone).
K ATAKE , K. (1961). The strength for tension and bursting of human fasciae. J. Kyoto Pref. Med. Univ. 69,
484–488.
K AZARIAN , L.E. (1982). Injuries to the human spinal column: Biomechanics and injury classification. Exerc.
Sport Sci. Rev. 9, 297–352.
K AZARIAN , L.E., B EERS , K., H ERNANDEZ , J. (1979). Spinal injuries in the F/FB-111 crew escape system.
Aviat. Space Environ. Med. 50, 948–957.
K EITHEL , L.M. (1972). Deformation of the thoracolumbar intervertebral joints in response to external loads.
J. Bone Joint Surgery A 54 (3).
K IMPARA, H., I WAMOTO, M., M IKI , K. (2002). Development of a small female FEM model. In: Proc. JSAE
Spring Congress, No. 59-02, Paper No. 20025242, pp. 1–4 (in Japanese).
K ING , A.I. (1993). Injury to the thoraco-lumbar spine and pelvis. In: Nahum, A.M., Melvin, J.W. (eds.),
K IRSCH , R.F., B OSKOV , D., RYMER , W.Z. (1994). Muscle stiffness during transient and continuous move-
ments of cat muscle: Perturbation characteristics and physiological relevance. IEEE Trans. Biomedical
Engrg. 41 (8), 758–770.
K ISIELEWICZ, T.K., A NDOH, K. (1994). Critical issues in biomechanical tests and simulations of impact
injury. In: PAM – User’s Conference in Asia, PUCA ’94, Shin-Yokohama, ESI Software S.A., 99, rue des
442 E. Haug et al.
Solets, BP 80112, 94513 Rungis Cedex, France, pp. 209–222, and: World Congress on Computational
Mechanics, WCCM ’94, Makuhari, Japan, August 1994.
K LEINBERGER, M., S UN, E., E PPINGER, R.H., K UPPA, S., S AUL, R. (1998). Development of improved
injury criteria for the assessment of advanced automotive restraint systems. NHTSA report.
K RESS , T.A., S NIDER , J.N., P ORTA , D.J., F ULLER , P.M., WASSERMAN , J.F., T UCKER , G.V. (1993). Hu-
man femur response to impact loading. In: Int. IRCOBI Conf. on the Biomechanics of Trauma, Eindhoven,
the Netherlands. IRCOBI Secretariat, Bron, France, pp. 93–104.
K ROELL , C.K. (1971). Thoracic response to blunt frontal loading. In: Backaitis, S.H. (ed.), Biomechanics of
Impact Injury and Injury Tolerances of the Thorax-Shoulder Complex (Society of Automotive Engineers),
pp. 51–80. Paper No. PT-45.
K ROELL , C.K., A LLEN , S.D., WARNER , C.Y., P ERL , T.R. (1986). Interrelationship of velocity and chest
compression in blunt thoracic impact to swine II. In: Proc. 30th Stapp Car Crash Conference, pp. 99–121.
Paper No. 861881.
K ROELL , C.K., S CHNEIDER , D.C., N AHUM , A.M. (1971). Impact tolerance and response of the human
thorax. In: Proc. 15th Stapp Car Crash Conference, pp. 84–134. Paper No. 710851.
K ROELL , C.K., S CHNEIDER , D.C., N AHUM , A.M. (1974). Impact tolerance and response of the human
thorax II. In: Proc. 18th Stapp Car Crash Conference, pp. 383–457. Paper No. 741187.
K ROONENBERG , A. VAN DEN, T HUNNISSEN , J., W ISMANS , J. (1997). A human model for low severity
rear-impacts. In: IRCOBI Conference.
K RYLOW , A.M., S ANDEROCK , T.G. (1996). Dynamic force responses of muscle involving eccentric con-
traction. J. Biomech. 30 (1), 27–33.
K ULAK , R.F., B ELYTSCHKO , T.B., S CHULTZ , A.B., G ALANTE , J.O. (1976). Nonlinear behavior of the
human intervertebral disc under axial load. J. Biomech. 9.
L ANIR , Y., F UNG , Y.C. (1974). Two-dimensional mechanical properties of rabbit skin – II. Experiment
results. J. Biomech. 7, 171–182.
L ASKY , I.I., S IEGEL , A.W., N AHUM , A.M. (1968). Automotive cardio-thoracic injuries: A medical-
engineering analysis. In: Automotive Engineering Congress, Detroit, MI. Paper No. 680052.
L EE , S-H., C HOI , H-Y. (2000). Finite element modeling of human head–neck complex for crashworthiness
simulation. Hongik J. Sci. Technol. 4, 1–15.
L EE , M.C., H AUT , R.C. (1989). Insensitivity of tensile failure properties of human bridging veins to strain
rate: implications in biomechanics of subdural hematoma. J. Biomech. 22, 537–542.
L EE , M.C., M ELVIN , J.W., U ENO , K. (1987). Finite element analysis of traumatic subdural hematoma. In:
L EE , J.B., YANG , K.H. (2001). Development of a finite element model of the human abdomen. In: Proc.
45th Stapp Car Crash Conference, pp. 1–22. Paper No. 2001-22-0004.
L ESTINA , D.C., K UHLMANN , T.P., K EATS , T.E., M AXWELL A LLEY , R. (1992). Mechanism of fracture in
ankle and foot injuries to drivers in motor vehicles. In: Proc. 36th Stapp Car Crash Conference, pp. 59–68.
Paper No. 922515.
L EVINE , R. (1993). Injury to the extremities. In: Nahum, A.M., Melvin, J.W. (eds.), Accidental Injury –
Biomechanics and Prevention (Springer, Berlin), pp. 460–491 (Chapter 18).
L EWIS , G., S HAW , K.M. (1997a). Modeling the tensile behavior of human achilles tendon. Biomed. Mater.
Engrg. 7 (4), 231–244.
L EWIS , G., S HAW , K.M. (1997b). Tensile properties of human tendo achillis: Effect of donor age and strain
rate. J. Foot Ankle Surgery 36 (6), 435–445.
L IN , H.S., L IU , Y.K., R AY , G., N IKRAVESH , P. (1978). Mechanical response of the lumbar intervertebral
joint under physiological (complex) loading. J. Bone Joint Surgery A 60 (1), 41–55.
L INDAHL , O. (1976). Mechanical properties of dried defatted spongy bone. Acta Orthop. Scand. 47, 11–19.
L INDE , F., et al. (1989). Energy absorptive properties of human trabecular bone specimens during axial
compression. J. Orth. Res. 7, 432–439.
L ISSNER , H.R., L EBOW , M., E VANS , F.G. (1960). Experimental studies on the relation between acceleration
and intracranial pressure changes in man. Surgery Gynecol. Obstet. 111, 329–338.
L IZEE , E., ROBIN , S., S ONG , E., B ERTHOLON , N., L E C OZ , J.Y., B ESNAULT , B., L AVASTE , F. (1998).
Development of a 3D finite element model of the human body. In: Proc. 42nd Stapp Car Crash Conference,
pp. 1–23. Paper No. 983152.
References 443
L IZEE , E., S ONG , E., et al. (1998). Finite element model of the human thorax validated in frontal, oblique
and lateral impacts: A tool to evaluate new restraint systems. In: Proceedings of the 1998 International
IRCOBI Conference.
L ÖHNER , R. (1990). Three-dimensional fluid-structure Interaction using a finite element solver and adaptive
re-meshing. Comput. Syst. Engrg. 1 (2–4), 257–272.
L OWENHIELM , P. (1974). Dynamic properties of parasagittal bridging veins. Z. Rechtsmedizin 74, 55–62.
L UNDBERG , A., G OLDIE , I., K ALIN , B., S ELVIK , G. (1989). Kinematics of the ankle/foot complex: Plan-
tarflexion and dorsiflexion. Foot & Ankle 9 (4), 194–200.
L UNDBERG , A., S VENSSON , O., B YLUND , C., G OLDIE , I., S ELVIK , G. (1989). Kinematics of the an-
kle/foot complex Part 2: Pronation and supination. Foot & Ankle 9 (5), 248–253.
M A , S., Z AHALAK , G.I. (1991). A distribution-moment model of energetics in skeletal muscle. J. Bio-
mech. 24 (1), 21–35.
M A , D., O BERGEFELL , L.A., R IZER , L.A. (1995). Development of human articulating joint model parame-
ters for crash dynamics simulations. In: Proc. 39th Stapp Car Crash Conference, pp. 239–250. Paper No.
952726.
M AENO , T., H ASEGAWA , J. (2001). Development of a finite element model of the total human model for
safety (THUMS) and application to car-pedestrian impacts. In: 17th International ESV Conference. Paper
No. 494.
M AKHSOUS , M., H ÖGFORS , C., S IEMIEN ’ SKI , A., P ETERSON , B. (1999). Total shoulder and relative mus-
cle strength in the scapular plane. J. Biomech. 32, 1213–1220.
M ANSCHOT , J.F., B RAKKEE , A.J. (1986a). The measurement and modelling of the mechanical properties
of human skin in vivo – I. The measurement. J. Biomech. 19 (7), 511–515.
M ANSCHOT , J.F., B RAKKEE , A.J. (1986b). The measurement and modelling of the mechanical properties
of human skin in vivo – II. The model. J. Biomech. 19 (7), 517–521.
M ARGULIES , S.S., T HIBAULT , L.E. (1989). An analytic model of traumatic diffuse brain injury. J. Biomech.
Engrg. 111, 241–249.
M ARGULIES , S.S., T HIBAULT , L.E. (1992). A proposed tolerance criterion for diffuse axonal injury in man.
J. Biomech. 25 (8), 917–923.
M ARGULIES , S.S., T HIBAULT , L.E., G ENNARELLI , T.A. (1990). Physical model simulations of brain injury
in the primate. J. Biomech. 23, 823–836.
M ARKOLF , K.L. (1972). Deformation of the thoracolumbar intervertebral joints in response to external loads.
J. Bone Joint Surgery A 54 (3).
M ARKOLF , K.L., M ORRIS , J.M. (1974). The structural components of the intervertebral disc. J. Bone Joint
Surgery A 56 (4).
M ARTENS , M., et al. (1983). The mechanical characteristics of cancellous bone at the upper femoral region.
J. Biomech. 16, 971–983.
M ARTIN , J., T HOMPSON , G. (1986). Achilles tendon rupture. CORR 210, 216–218.
M ARTINI , F.H., T IMMONS , M.J., TALLITSCH , B. (2003). Human Anatomy, fourth ed. (Pearson Education,
Inc., publishing as Benjamin Cummings, Upper Saddle River, NJ), ISBN 0-13-061569-2.
M ASSON , C., C ESARI , D., BASILE , F., B EAUGONIN , M., T RAMECON , A., A LLAIN , J.C., H AUG , E.
(1999). Quasi static ankle/foot complex behavior: Experimental tests and numerical simulations. In: IR-
COBI Conference, Spain.
M AUREL, W. (1998). 3D modelling of the human upper limb including the biomechanics of joints, muscles
and soft tissues, Thèse N◦ 1906, Ecole Polytechnique Federale de Lausanne.
M C C ALDEN , R.W., et al. (1993). Age-related in the tensile properties of cortical bone. J. Bone Joint Surgery
A 75, 1193–1205.
M C C LURE , P., S IEGLER , S., N OBILINI , R. (1998). Three-dimensional flexibility characteristics of the hu-
man cervical spine in vivo. Spine 23 (2), 216–223.
M C E LHANEY , J.H. (1966). Dynamic response of bone and muscle tissue. J. Appl. Physiol., 1231.
M C E LHANEY , J.H., D OHERTY , B.J., PAVER , J.G., M YERS , B.S., G RAY , L. (1988). Combined bending
and axial loading responses of the human cervical spine. In: Proc. 32nd Stapp Car Crash Conference,
pp. 21–28. Paper No. 881709.
444 E. Haug et al.
M C E LHANEY , J.H., M ELVIN , J.W., ROBERTS , V.L., P ORTNOY , H.D. (1973). Dynamic characteristics of
the tissues of the head. In: Kennedi, R.M. (ed.), Perspectives in Biomedical Engineering (MacMillan,
London), pp. 215–222.
M C E LHANEY , J.H., M YERS , B.S. (1993). Biomechanical aspects of cervical trauma. In: Nahum, A.M.,
Melvin, J.W. (eds.), Accidental Injury – Biomechanics and Prevention (Springer, Berlin), pp. 311–361
(Chapter 14).
M C E LHANEY , J.H., PAVER , J.G., M C C RACKIN , H.J., M AXWELL , G.M. (1983). Cervical spine compres-
sion responses. In: Proc. 27th Stapp Car Crash Conference, pp. 163–177. Paper No. 831615.
M C E LHANEY , D.A., et al. (1970). Mechanical properties of cranial bone. J. Biomech. 3, 495–511.
M EANY , D.F., S MITH , D., ROSS , T., G ENNARELLI , T.A. (1993). Diffuse axonal injury in the miniature
pig: Biomechanical development and injury threshold. ASME Crashworthiness and Occupant Protection
Systems 25, 169–175.
M ELVIN , J.W., L IGHTHALL , J.W., U ENO , K. (1993). Brain injury biomechanics. In: Nahum, A.M., Melvin,
J.W. (eds.), Accidental Injury – Biomechanics and Prevention (Springer, Berlin), pp. 268–291 (Chap-
ter 12).
M ELVIN , J.W., M C E LHANEY , J.H., ROBERTS , V.L. (1970). Development of mechanical model of human
head – determination of tissue properties and synthetic substitute materials. In: Proc. 14th Stapp Car Crash
M ELVIN , J.W., S TALNAKER , R.L., ROBERTS , V.L. (1973). Impact injury mechanisms in abdominal organs.
In: Proc. 17th Stapp Car Crash Conference Proceedings, pp. 115–126. Paper No. 730968.
M ERTZ , H.J., PATRIC , L.M. (1971). Strength and response of human neck. In: Proc. 15th Stapp Car Crash
M ERTZ, H.J. (1984). A procedure for normalizing impact response data. SAE Paper 840884, Washington,
DC.
M ERTZ , H.J. (1993). Anthropomorphic test devices. In: Nahum, A.M., Melvin, J.W. (eds.), Accidental Injury
– Biomechanics and Prevention (Springer, Berlin), pp. 66–84 (Chapter 4).
M ILLER , K. (2000). Constitutive modeling of abdominal organs. J. Biomech. 33, 367–373.
M ILLER , K., C HINZEI , K. (1997). Constitutive modelling of brain tissue: Experiment and theory. J. Bio-
mech. 30, 1115–1121.
M ILLER , J.A.A., S CHULTZ , A.B., WARWICK , D.N., S PENCER , D.L. (1986). Mechanical properties for
lumbar spine motion segments under large loads. J. Biomech. 19 (1), 79–84.
M OFFATT , C.A., A DVANI , S.H., L IN , C.J. (1971). Analytical end experimental investigations of human
spine flexure. Amer. Soc. Mech. Engrg. 71-WA/BHF-7.
M OMERSTEEG , T.J.A., B LASKEVOORT , L., H UISKES , R., K OOLOOS , J.G., K AUER , J.M.G. (1996). Char-
acterization of the mechanical behavior of human knee ligaments: A numerical–experimental approach.
J. Biomech. 29 (2), 151–160.
M ONAGHAN , J.J. (1988). An introduction into SPH. Comput. Phys. Comm. 48, 89–96.
M ONAGHAN , J.J., G INGOLD , R.A. (1983). Shock simulation by the particle method SPH. Comput. Phys. 52,
374–398.
M ORGAN , D.L. (1990). New insights into the behavior of muscle during active lengthening. Biophys. J. 57,
209–221.
M ORGAN , R.M., E PPINGER , R.H., H ENNESSEY , B. (1991). Ankle joint injury mechanism for adults in
frontal automotive impact. In: Proc. 35th Stapp Car Crash Conference Proceedings, pp. 189–198. Paper
No. 912902.
M ORONEY , S.P., S CHULTZ , A.B., M ILLER , J.A.A. (1988). Analysis and measurement of neck load. J. Or-
thopaedic Res. 6, 713–720.
M ORONEY , S.P., S CHULTZ , A.B., M ILLER , J.A.A., A NDERSSON , G.B.J. (1988). Load-displacement prop-
erties of lower cervical spine motion segments. J. Biomech. 21 (9), 769–779.
M OSEKILDE , L., et al. (1986). Normal vertebral body size and compressive strength: Relations to age and to
vertebral and iliac trabecular bone compressive strength. Bone 7, 207–212.
M OW , V.C., H AYES , W.C. (1991). Basic Orthopaedic Biomechanics (Raven Press, New York).
M YERS , B.S., M C E LHANEY , J.H., D OHERTY , B.J., PAVER , J.G., G RAY , L. (1991). The role of torsion in
cervical spinal injury. Spine 16 (8), 870–874.
References 445
M YERS , B.S., M C E LHANEY , J.H., R ICHARDSON , W.J., N IGHTINGALE , R., D OHERTY , B.J. (1991). The
influence of end condition on human cervical spine injury mechanisms. In: Proc. 35th Stapp Car Crash
M YERS , B.S., VAN E E , C.A., C AMACHO , D.L.A., W OOLLEY , C.T., B EST , T.M. (1995). On the structural
properties of mammalian skeletal muscle and its relevance to human cervical impact dynamics. In: Proc.
39th Stapp Car Crash Conference, pp. 203–214. Paper No. 952723.
M YKLEBUST , J.B., P INTAR , F. (1988). Tensile strength of spinal ligaments. J. Spine 13 (5).
N AGASAKA , K., I WAMOTO , M., M IZUNO , K., M IKI , K., H ASEGAWA , J. (2002). Pedestrian injury analy-
sis using the human FE model (Part I, development and validation of the lower extremity model and
application to damage evaluation of knee ligaments). In: Proc. JSME 14th Bioengineering Conference,
pp. 141–142 (in Japanese).
N AHUM , M., G ADD , C.W., S CHNEIDER , D.C., K ROELL , C.K. (1970). Deflection of the human tho-
rax under sternal impact. In: Proceedings of the International Automobile Safety Conference, Detroit.
SAE 700400.
N AHUM , A.M., M ELVIN , J.W. (eds.) (1993). Accidental Injury: Biomechanics and Prevention (Springer,
New York).
N AHUM , M., S CHNEIDER , D.C., K ROELL , C.K. (1975). Cadaver skeletal response to blunt thoracic impact.
N AHUM , A.M., S IEGEL , A.W., H IGHT , P.V., B ROOKS , S.H. (1968). Lower extremity injuries of front seat
occupants. In: Proc. 11th Stapp Car Crash Conference Proceedings. Paper No. 680483.
N AHUM , A.M., S MITH , R., WARD , C.C. (1977). Intracranial pressure dynamics during head impact. In:
N EUMANN , P., K ELLER , T.S., E KSTROM , L., P ERRY , L., H ANSSON , T.H., S PENGLER , D.M. (1992).
Mechanical properties of the human lumbar anterior longitudinal ligament. J. Biomech. 25 (10), 1185–
1194.
N EWMAN , J.A. (1993). Biomechanics of head trauma: Head protection. In: Nahum, A.M., Melvin, J.W.
(eds.), Accidental Injury – Biomechanics and Prevention (Springer, Berlin), pp. 292–310 (Chapter 13).
NHTSA/CIREN (1997). Crash test results from William Lehman Injury Research Center at the Ry-
der Trauma Center. In: First Annual CIREN Conference, October 20, 1997, http://www-nrd.nhtsa.dot.
gov/departments/nrd-50/ciren/ciren1.html.
N ICOLL , E.A. (1949). Fractures of the dorso-lumbar spine. J. Bone Joint Surgery B 31, 376–393.
N IGG , B.M., S KARVAN , G., F RANK , C.B. (1990). Elongation and forces of ankle ligaments in a physiolog-
ical range of motion. Foot & Ankle 11 (1), 30–40.
N IGHTINGALE , R.W., W INKELSTEIN , B.A., K NAUB , K.E., R ICHARDSON , W.J., L UCK , J.F., M YERS ,
B.S. (2002). Comparative strengths and structural properties of the upper and lower cervical spine in
flexion and extension. J. Biomech. 35, 725–732.
N ITSCHE, S., H AUG, E., K ISIELEWICZ, L.T. (1996). Validation of a finite element model of the human neck.
In: PUCA ’96, Nihon ESI, Japan, ESI Software S.A., 99, rue des Solets, BP 80112, 94513 Rungis Cedex,
France, pp. 203–220.
N ORDIN , M., F RANKEL , V.H. (1989). Basic Biomechanics of the Musculoskeletal System, second ed. (Lea
& Febiger, Philadelphia, PA).
N OVOTNY , J.E. (1993). Spinal biomechanics. J. Biomech. Engrg. 115.
N OYES , F.R., G ROOD , E.S. (1978). The strength of the anterior cruciate ligaments in humans and rhesus
monkey. J. Bone Joint Surgery A 58 (68), 1074–1082.
N USHOLTZ , G.S., W ILEY , B., G LASCOE , L.G. (1995). Cavitation/boundary effects in a simple head impact
model. Aviation Space and Environmental Medicine 66 (7), 661–667.
N YQUIST , G.W., C HENG , R., E L -B OHY , A.A.R., K ING , A.I. (1985). Tibia bending: Strength and response.
In: Proc. 29th STAPP Car Crash Conference, pp. 99–112. Paper No. 851728.
O MMAYA , A.K. (1967). Mechanical properties of tissues of the nervous system. J. Biomech. 1, 127–138.
O MMAYA , A.K., G ENNARELLI , T.A. (1974). Cerebral concussion and traumatic unconsciousness: Correla-
tion of experimental and clinical observations on blunt head injuries. Brain 97, 633–654.
O MMAYA , A.K., H IRSCH , A.E. (1971). Tolerances for cerebral concussion from head impact and whiplash
in primates. J. Biomech. 4, 13–31.
446 E. Haug et al.
O MMAYA , A.K., H IRSCH , A.E., F LAMM , E.S., M AHONE , R.H. (1966). Cerebral concussion in the monkey:
An experimental model. Science 153, 211–212.
O NO , K., K ANEOKA , K., S UN , E.A., TAKHOUNTS , E.G., E PPINGER , R.H. (2001). Biomechanical re-
sponse of human cervical spine to direct loading of the head. In: IRCOBI Conference.
O NO , K., K ANEOKA , K., W ITTEK , A., K AJZER , J. (1997). Cervical injury mechanism based on the analysis
of human cervical vertebral motion and head-neck-torso kinematics during low speed rear impacts. In:
O NO , K., K IKUCHI , A., N AKAMURA , M., K OBAYASHI , H., N AKAMURA , N. (1980). Human head tolerance
to sagittal impact reliable estimation deduced from experimental head injury using subhuman primates and
human cadaver skulls. In: Proc. 24th Stapp Car Crash Conference, pp. 101–160. Paper No. 801303.
O NO , K., et al. (1999). Relationship between localized spine deformation and cervical vertebral motions for
low speed rear impacts using human volunteers. In: IRCOBI Conference, Spain, pp. 149–164.
O SHITA , F., O MORI , K., N AKAHIRA , Y., M IKI , K. (2002). Development of a finite element model of the
human body. In: Proc. 7th International LS-DYNA Users Conference, Detroit, pp. 3-37–3-48.
O TTE , D., VON R HEINBABEN , H., Z WIPP , H. (1992). Biomechanics of injuries to the foot and ankle joint
of car drivers and improvements for an optimal car floor development. In: Proc. 36th Stapp Car Crash
Conference Proceedings, pp. 43–58. Paper No. 922514.
O TTENSMEYER , M.P., S ALISBURY , J. (2001). In vivo data acquisition instrument for solid organ mechanical
property measurement. In: Niessen, W., Viergever, M. (eds.), MICCAI 2001. In: Lecture Notes in Comput.
Sci. 2208 (Springer, Berlin), pp. 975–982.
O XLUND , H., A NDREASSON , T.T. (1980). The role of hyaluronic acid, collagen and elastin in the mechanical
properties of connective tissues. J. Anat. 131, 611–620.
PALANIAPPAN , P. J R ., W IPASURAMONTON , P., B EGEMAN , P., TANAVDE , A.S., Z HU , F.A. (1999). Three
dimensional finite element model of the human arm. In: Proc. 43rd Stapp Car Crash Conference, pp. 351–
363. Paper No. 99SC25.
PANJABI , M.M., B RAND , R.A. (1976). Mechanical properties of the human thoracic spine as shown by
three-dimensional load displacement curves. J. Bone Joint Surgery A 58 (5).
PANJABI , M.M., C RISCO , J.J., VASAVADA , A., O DA , T., C HOLEWICKI , J., N IBU , K., S HIN , E. (2001). Me-
chanical properties of the human cervical spine as shown by three-dimensional load-displacement curves.
Spine 26 (24), 2692–2700.
PANJABI , M.M., D VORAK , J., D URANCEAU , J., et al. (1988). Three-dimensional movements of the upper
cervical spine. J. Spine 13 (7), 726.
PANJABI , M., J ORNEUS , L., G REENSTEIN , G. (1984). Physical properties of lumbar spine ligaments. Trans.
Orthop. Res. Soc. 9, 112.
PARENTEAU , C.S., V IANO , D.C. (1996). Kinematics study of the ankle-subtalar joints. J. Biomech. Engrg.,
Ph.D. Thesis work.
PARENTEAU , C.S., V IANO , D.C., P ETIT , P.Y. (1996). Biomechanical properties of ankle-subtalar joints in
quasi-static loading to failure. J. Biomech. Engrg.
PATHRIA , M.N., R ESNIK , D. (1993). Radiologic analysis of trauma. In: Nahum, A.M., Melvin, J.W (eds.),
PATTIMORE , D., WARD , E., T HOMAS , P., B RADFORD , M. (1991). The nature and cause of lower limb
injuries in car crashes. In: Proc. 35th Stapp Car Crash Conference, pp. 177–188. Paper No. 912901.
P EINDL , R.D., E NGIN , A.E. (1987). On the biomechanics of human shoulder complex – II. Passive resistive
properties beyond the shoulder complex sinus. J. Biomech. 20 (2), 118–134.
P ENN , R.D., C LASEN , R.A. (1982). Traumatic brain swelling and edema. In: Cooper, P.R. (ed.), Head Injury
(Williams and Wilkins, Baltimore), pp. 233–256.
P ENNING , L. (1979). Normal movements of the cervical spine. Amer. J. Roentgenol., 130–317.
P ENNING , L., W ILMARK , J.T. (1987). Rotation of the cervical spine. J. Spine 12 (8), 732.
P IKE , J.A. (1990). Automotive Safety (Society of Automotive Engineers). ISBN 1-56091-007-0.
P INTAR , F.A., Y OGANANDAN , N., E PPINGER , R.H. (1998). Response and tolerance of the human forearm
to impact loading. In: Proc. 42nd Stapp Car Crash Conference, pp. 1–8. Paper No. 983149.
P LANK , G.R., K LEINBERGER , M., E PPINGER , R.H. (1994). Finite element modeling and analysis of tho-
rax/restraint system interaction. In: The 14th ESV International Technical Conference on the Enhanced
Safety of Vehicles (Munich, Germany).
References 447
P OPE , M.E., K ROELL , C.K., V IANO , D.C., WARNER , C.Y., A LLEN , S.D. (1979). Postural influences on
thoracic Impact. In: Proc. 23rd Stapp Car Crash Conference, pp. 765–795. Paper No. 791028.
P ORTIER , L., T ROSSEILLE , X., L E C OZ , J.-Y., L AVASTE , F., C OLTAT , J.-C. (1993). Lower leg injuries in
real-world frontal accidents. In: 28th International IRCOBI Conference on the Biomechanics of Trauma,
pp. 57–78.
P RADAS , M.M., C ALLEJA , R.D. (1990). Nonlinear viscoelastic behaviour of the flexor tendon of the human
hand. J. Biomech. 23 (8), 773–781.
P RASAD , P. (1990). Comparative evaluation of the dynamic responses of the Hybrid II and the Hybrid III
dummies. In: Proc. 34th Stapp Car Crash Conference, pp. 175–183. Paper No. 902318.
P RASAD , P., C HOU , C.C. (1993). A review of mathematical occupant simulation models. In: Nahum, A.M.,
Melvin, J.W. (eds.), Accidental Injury – Biomechanics and Prevention (Springer, Berlin), pp. 102–150
(Chapter 6).
P RASAD , P., K ING , A.I. (1994). An experimentally validated dynamic model of the spine. J. Appl. Mech.,
546–550.
P RING , D.J., A MIS , A.A., C OOMBS , R.R. (1985). The mechanical properties of human flexor tendons in
relation to artificial tendons. J. Hand Surgery [Br] 10 (3), 331–336.
P UTZ , R., PABST , R. (2000). Sobotta – Atlas der Anatomie des Menschen, twenty-first ed. (Urban&Fischer,
München, Jena).
Q UHAN, et al. (1989). Comparison of trabecular and cortical tissue moduli from human iliac crests. J. Orthop.
Res. 7, 876–884.
R ACK , P.M.H., W ESTBURY , D.R. (1969). The effect of length and stimulus rate on tension in the isometric
cat soleus muscle. J. Physiol. 204, 443–460.
R EILLY , D.T., B URSTEIN , A.H., F RANKEL , V.H. (1974). The elastic modulus for bone. J. Biomech. 7,
271–275.
R EILLY , D.T., et al. (1975). The elastic and ultimate properties of compact bone tissue. J. Biomech. 8, 395–
405.
R ENAUDIN , F., G UILLEMOT , H., P ÉCHEUX , C., L ESAGE , F., L AVASTE , F., S KALLI , W. (1993). A 3D finite
element model of the pelvis in side impact. In: Proc. 37th Stapp Car Crash Conference, pp. 241–252. Paper
No. 933130.
R IETBERGEN, B. VAN (1996). Mechanical behavior and adaptation of trabecular bone in relation to bone
morphology. PhD Thesis Nijmegen University. ISBN 90-9010006-7.
R IETBERGEN , B. VAN, M ÜLLER , R., U LRICH , D., RÜEGSEGGER , P., H UISKES , R. (1998). Tissue stresses
and strain in trabeculae of a canine proximal femur can be quantified from computer reconstructions.
J. Biomech.
ROAF , R. (1960). A study of the mechanics of spinal injury. J. Bone Joint Surgery B 42, 810–823.
ROBBINS , D.H. (1983). Anthropometry of Motor Vehicle Occupants, vol. 2: Mid-sized Male, vol. 3: Small
Female and Large Male (UMTRI-83-53-2, University of Michigan).
ROBBINS , D.H., S CHNEIDER , L.W., S NYDER , R.G., P FLUG , M., H AFFNER , M. (1983). Seated posture of
vehicle occupants. In: Proc. 32nd Stapp Car Crash Conference, pp. 199–224. Paper No. 831617.
ROBBY 1 (1997). User’s Guides, ESI Software S.A., 99, rue des Solets, BP 80112, 94513 Rungis Cedex,
France.
ROBBY 2 (1998). User’s Guides, ESI Software S.A., 99, rue des Solets, BP 80112, 94513 Rungis Cedex,
France.
ROBIN, S. (1999). Validation data base, Report PSA/990331/T0/DA, LAB PSA Peugeot Citroën – Renault,
132 Rue des Suisses, F-92000 Nanterre.
ROBIN , S. (2001). HUMOS: Human model for safety – A joint effort towards the development of refined
human-like car occupant models. In: 17th ESV Conference. Paper Number 297.
ROBINA (1998). Internal draft report, ESI Software S.A., 99, rue des Solets, BP 80112, 94513 Rungis Cedex,
France.
ROCKOFF , S.D., et al. (1969). The relative contribution of trabecular and cortical bone to the strength of
human lumbar vertebrae. Calc. Tiss. Res. 3, 163.
ROHEN , J.W., Y OKOCHI , C. (1983). In: Color Atlas of Anatomy (IGAKU–SHOIN Medical Publishers),
p. 349.
448 E. Haug et al.
ROSE , J.L., G ORDON , S.L., M OSKOWITZ , G. (1974). Dynamic photoelastic model analysis of impact to the
human skull. J. Biomech. 7, 193–199.
ROUHANA , S.W. (1993). Biomechanics of abdominal trauma. In: Nahum, A.M., Melvin, J.W. (eds.), Acci-
dental Injury – Biomechanics and Prevention (Springer, Berlin), pp. 391–428 (Chapter 16).
RUAN , J.S., K HALIL , T.B., K ING , A.I. (1991). Human head dynamic response to side impact by finite
element modeling. J. Biomech. Engrg. 113, 267–283. PhD Dissertation, Wayne State University.
RUAN , J.S., K HALIL , T.B., K ING , A.I. (1993). Finite element modeling of direct head impact. In: Proc.
RUAN , J.S., K HALIL , T.B., K ING , A.I. (1994). Dynamic response of the human head to impact by three-
dimensional finite element analysis. ASME J. Biomech. Engrg. 116, 44–50.
RUAN , J.S., P RASAD , P. (1994). Head injury assessment in frontal impacts by mathematical modeling. In:
Proc. 38th Stapp Car Crash Conference, pp. 111–121. Paper No. 942212.
RUDOLF , C., F ELLHAUER, A., S CHAUB, S., M ARCA, C., B EAUGONIN, M. (2002). OOP Simulation with a
5th Percentile Deformable Human Model. In: EuroPam 2002, ESI Software S.A., 99, rue des Solets, BP
80112, 94513 Rungis Cedex, France.
SAE E NGINEERING A ID 23 (1986). User’s Manual for the 50th Percentile HYBRID III Test Dummy.
S ANCES J R , A., et al. (1982). Head and spine injuries. In: AGARD Conference on Injury Mechanisms, Pre-
vention and Cost, Köln, Germany, pp. 13-1–13-33.
S ACRESTE , J., B RUN -C ASSAN , F., FAYON , A., TARRIERE , C., G OT , C., PATEL , A. (1982). Proposal for
a thorax tolerance level in side impacts based on 62 tests performed with cadavers having known bone
condition. In: Proc. 26th Stapp Car Crash Conference, pp. 155–171. Paper No. 821157.
S CHMIDT, G., K ALLIERIS , D., BARZ, J., M ATTERN, R., S CHULZ, F., S CHÜLER, F. (1978). Biomechanics –
Determination of the mechanical loadability limits of the occupants of a motor vehicle. Final report at the
end of project (31-12-1978) – Research project, No. 3906, Institute of Forensic Medicine, Heidelberg.
S CHNECK , D.J. (1992). Mechanics of Muscle (New York University Press, New York).
S CHNEIDER, L.W., K ING, A.I., B EEBE, M.S. (1990). Design requirements and specifications, Thorax ab-
domen development task. Interim report: Trauma assessment device development program. Report No.
DOT-HS-807-511.
S CHOENFELD , C.M., et al. (1974). Mechanical properties of human cancellous bone in the femoral head.
Med. Biol. Engrg. 12, 313–317.
S COTT , W.E. (1981). Epidemiology of head and neck trauma in victims of motor vehicle accidents. Head
and Neck Criteria. In: Ommaya, A.K. (ed.), A Consensus Workshop (US Department of Transportation,
National Highway Traffic Safety Administration, Washington, DC), pp. 3–6.
S EDLIN , E.D., et al. (1965). A rheological model for cortical bone. In: A Study of the Physical Properties of
Human Femoral Samples. In: Acta Orthop Scand. Supplementum 83.
S EDLIN , E.D., et al. (1966). Factors affecting the determination of the physical properties of femoral cortical
bone. Acta Orthop. Scand. 37, 29–48.
S EIREG , A., A RVIKAR , R. (1989). Biomechanical Analysis of the Musculoskeletal Structure for Medicine
and Sports (Hemisphere, Washington, DC).
S HAH , C.S., YANG , K.H., H ARDY , W.N., WANG , H.K., K ING , A.I. (2001). Development of a computer
model to predict aortic rupture due to impact loading. Stapp Car Crash J. 45, 161–182.
S HELTON , F.E., B UTLER , D.L., F EDER , S.M. (1993). Shear stress transmission in the patellar tendon and
mediacollateral ligament. Adv. Bioengrg. BED ASME 26, 271.
S HUCK , L.Z., A DVANI , S.H. (1972). Rheological response of human brain tissue in shear. ASME J. Basic
Engrg., 905–911.
S HUCK, L.Z., H AYNES , R.R., F OGLE, J.L. (1970). Determination of viscoelastic properties of human brain
tissue, ASME Paper N◦ 70-BHF-12.
S HUGAR , T.A. (1975). Transient structural response of the linear skull-brain system. In: Proc. 19th Stapp
Car Crash Conference, pp. 581–614. Paper No. 751161.
S KAGGS , D.L., W EIDENBAUM , M., I ATRIDIS , J.C., R ATCLIFFE , A., M OW , V.C. (1994). Regional varia-
tions in tensile properties and biomechanical composition of the human lumbar annulus fibrosus. Spine 19,
1310–1319.
S ONNERUP , L. (1972). A semi-experimental stress analysis of the human intervertebral disk in compression.
Experimental Mech.
References 449
S ONODA , T. (1962). Studies on the strength for compression, tension and torsion of the human vertebral
column. J. Kyoto Pref. Med. Univ. 71, 659–702.
S PILKER , R.L., JAKOBS , D.M., S CHULTZ , A.B. (1986). Material constants for a finite element model of the
intervertebral disc with a fiber composite annulus. J. Biomech. Engrg. 108, 1–11.
S PITZER , V.M., W HITLOCK , D.G. (1998). Atlas of the Visible Human Male (Jones & Bartlett Publishers),
pp. 6, 7.
S TALNAKER , R.L., M C E LHANEY, J.H., ROBERTS , V.L., T ROLLOPE , L.L. (1973). Human torso response to
blunt trauma. In: Proceedings of the Symposium: Human Impact Response – Measurement and Simulation,
General Motors Research laboratories, New York (Plenum Press, London), pp. 181–199.
S TALNAKER , R.L., M OHAN , D. (1974). Human chest impact protection criteria. In: Proc. 3rd International
Conference on Occupant Protection (Society of Automotive Engineers, New York), pp. 384–393.
S TATES , J.D. (1986). Adult occupant injuries of the lower limb. In: Biomechanics and Medical Aspects of
Lower Limb Injuries. SAE 861927.
S TOCKIER , R.M., E PSTEIN , J.A., E PSTEIN , B.S. (1969). Seat belt trauma to the lumbar spine: An unusual
manifestation of the seat belt syndrome. J. Trauma 9, 508–513.
S TRUHL , S., et al. (1987). The distribution of mechanical properties of trabecular bone within vertebral bodies
and iliac crest: Correlation with computed tomography density. Trans. Orthop. Res. Soc., 262.
S TÜRTZ , G. (1980). Biomechanical data of children. In: Proc. 24th Stapp Car Crash Conference, pp. 513–
559. Paper No. 801313.
S VENSSON , M.Y., L ÖVSUND , P. (1992). A dummy for rear-end collisions – development and validation of
a new dummy-neck. In: Proceedings of IRCOBI Conference, Verona, Italy.
TADA , Y., N AGASHIMA , T. (1994). Modeling and simulation of brain lesions by the finite element method.
IEEE Engineering in Medicine and Biology 13 (4).
TARRIERE , C. (1981). Investigation of brain injuries using the CT scanner. In: Ommaya, A.K. (ed.), Head
and Neck Injury Criteria: A Consensus Work-Shop (US Department of Transportation, National Highway
Traffic Safety Administration, Washington, DC), pp. 39–49.
T ENNYSON , S.A., K ING , A.I. (1976a). A biodynamic model of the spinal column. In: Mathematical Mod-
eling Biodynamic Response to Impact, SAE Special Publication. SAE 760711.
T ENNYSON , S.A., K ING , A.I. (1976b). Electromyographic signals of the spinal musculature during +Gz
accelerations. In: 1976 Meeting of the International Society for the Study of the Lumbar Spine, Bermuda.
T HIBAULT , K.L., M ARGULIES , S.S. (1996). Material properties of the developing porcine brain. In: 1996
International IRCOBI Conference on the Biomechanics of Impact, Dublin, Ireland.
T HUNNISSEN , J., W ISMANS , J., E WING , C.L., T HOMAS , D.J. (1995). Human volunteer head–neck re-
sponse in frontal flexion: A new analysis. In: Proc. 39th Stapp Car Crash Conference, pp. 439–460. Paper
No. 952721.
T’ MER , S.T., E NGIN , A.E. (1989). Three-dimensional kinematic modelling of the human shoulder complex
– Part II: Mathematical modelling and solution via optimization. J. Biomech. Engrg. 111, 113–121.
T ORG , J.S. (ed.) (1982). Athletic Injuries to the Head, Neck and Face (Lea & Febiger, Philadelphia, PA).
T ORG , J.S., PAVLOV , H. (1991). Axial load “teardrop” fracture. In: Athletic Injuries to the Head, Neck and
Face, second ed. (Lea & Febiger, Philadelphia, PA).
T ROSSEILLE , X., TARRIERE , C., L AVASTE , F., G UILLON , F., D OMONT , A. (1992). Development of an
F.E.M. of the human head according to a specific test protocol. In: Proc. 36th Stapp Car Crash Conference,
pp. 235–253. Paper No. 922527.
T URQUIER , F., K ANG , H.S., T ROSSEILLE , X., W ILLINGER , R., L AVASTE , F., TARRIERE , C., D ÖMONT ,
A. (1996). Validation study of a 3D finite element head model against experimental data. In: Proc. 40th
U ENO , K., L IU , Y.K. (1987). A three-dimensional nonlinear finite element model of lumbar intervertebral
joint in torsion. J. Biomech. Engrg. 109, 200–209.
U ENO , K., M ELVIN , J.W., L I , L., L IGHTHALL , J.W. (1995). Development of tissue level brain injury criteria
by finite element analysis. J. Neurotrauma 12 (4).
U ENO , K., M ELVIN , J.W., L UNDQUIST , E., L EE , M.C. (1989). Two-dimensional finite element analysis of
human brain impact responses: Application of a scaling law. In: Crashworthiness and Occupant Protection
in Transportation Systems. In: AMD 106 (The American Society of Mechanical Engineers, New York),
pp. 123–124.
450 E. Haug et al.
U LRICH, D. (1998). Evaluation of the mechanical properties of bone with consideration of its microarchi-
tecture. Dissertation for the degree of doctor of the technical sciences of the Swiss Federal Institute of
Technology, Zürich, Switzerland.
VANCE , T.L., S OLOMONOV , M., BARATTA , R., Z EMBO , M., D’A MBROSIA , R.D. (1994). Comparison of
isometric and load moving length-tension models of two bicompartmental muscles. IEEE Trans. Biomed.
Engrg. 41 (8), 771–781.
V ERRIEST , J.-P., C HAPON , A. (1994). Validity of thoracic injury criteria based on the number of rib
fractures. In: Backaitis, S.H. (ed.), Biomechanics of Impact Injury and Injury Tolerances of the Thorax-
Shoulder Complex (Society of Automotive Engineers), pp. 719–727. SAE PT-45.
V IANO , D.C. (1986). Biomechanics of bone and tissue: A review of material properties and failure char-
acteristics. In: Symposium on Biomechanics and Medical Aspects of Lower Limb Injuries (San Diego,
California), pp. 33–63. SAE 861923.
V IANO , D.C. (1989). Biomechanical responses and injuries in blunt lateral impact. In: Proc. 33rd Stapp Car
Crash Conference, pp. 113–142. Paper No. 892432.
V IANO , D.C., L AU , V.K. (1983). Role of impact velocity and chest compression in thoracic injury. Aviat.
Space Environ. Med. 54 (1), 16–21.
V IIDIK , A. (1987). Properties of tendons and ligaments. In: Skalak, R., Chien, S. (eds.), Handbook of Bio-
engineering (McGraw-Hill, New York), pp. 6.1–6.19.
V IRGIN , W.J., L UDHIANA , P. (1951). Experimental investigations into the physical properties of the inter-
vertebral disc. J. Bone Joint Surgery B 33 (4).
V ISIBLE H UMAN P ROJECT (1994) (public release of male dataset on CD Rom), National Library of Medi-
cine. Visible Human Database. 8600 Rockville Pike, Bethseda, MD 20894.
V OIGT , M., B OJSEN -M OLLER , F., S IMONSEN , E.B., DYHRE -P OULSEN , P. (1995). The influence of tendon
Youngs modulus, dimensions and instantaneous moment arms on the efficiency of human movement.
J. Biomech. 28 (3), 281–291.
V OO , L., K UMARESAN , S., P INTAR , F.A., Y OGANANDAN , N., S ANCES J R , A. (1996). Finite element
models of the human head. Medical Biological Engrg. Comput. 34, 375–381.
WAINWRIGHT , S.A., B IGGS , W.D., C URREY, J.D. (1979). Mechanical Design in Organisms – Pliant Ma-
terials (Edward Arnold, London), pp. 110–143.
WALKE , A.E., K OLLROS , J.J., C ASE , T.J. (1944). The physiological basis of concussion. J. Neurosurg. 1,
103–116.
WANK , V., G UTEWORT, W. (1993). Modelling and simulation of muscular contractions with regard to
physiological parameters. In: Proc. of the XIV International Society of Biomechanics Congress, Paris
(International Society of Biomechanics), pp. 1450–1451.
WARD , C.C. (1982). Finite element models of the head and their use in brain injury research. In: Proc. 26th
WARD , C.C., T HOMSON , R.B. (1975). The development of a detailed finite element brain model. In: Proc.
WATANABE , I., I SHIHARA , T., F URUSU , K., K ATO , C., M IKI , K. (2001). Basic research of spleen FEM
model for impact analysis. In: Proc. of 2001 JSME Annual Congress, pp. 115–116 (in Japanese).
W EAVER , J.K., et al. (1966). Cancellous bone: Its strength and changes with aging and an evaluation of some
methods for measuring its mineral content. J. Bone Joint Surgery A 48, 289–299.
W IRHED , R. (1985). Anatomie et Science du Geste Sportif (Edition VIGOT). ISBN 2-7114-0944-9.
W HITE , A.A., PANJABI , M.M. (1978). Biomechanics of the Spine (Lippincott, Philadelphia, PA).
W HITE , A.A., PANJABI , M.M. (1990). Clinical biomechanics of the spine. J. Biomech., Lippincott Company,
second ed.
W ILLIAMS , J.L., et al. (1982). Properties and an anisotropic model of cancellous bone from the proximal
tibia epiphysis. J. Biomech. Engrg. 104, 50–56.
W ILLINGER , R., K ANG , H.S., D IAW , B. (1999). Three-dimensional human finite-element model validation
against two experimental impacts. Ann. Biomedical Engrg. 27, 403–410.
W ILSON -M AC D ONALD , J., W ILLIAMSON , D.M. (1988). Severe ligamentous injury of the ankle with rup-
tured tendo achillis and fractured neck talus. J. Trauma 28, 872–874.
W INTERS , J.M., S TARK , L. (1985). Analysis of fundamental human movement patterns through the use of
in-depth antagonistic muscle models. IEEE Trans. Biomedical Engrg. 12, 826–839.
References 451
W INTERS , J.M., S TARK , L. (1988). Estimated mechanical properties of synergistic muscles involved in
movements of a variety of human joints. J. Biomech. 21, 1027–1041.
W ISMANS , J.S., S PENNY , C. (1983). Performance requirements of mechanical necks in lateral flexion. In:
Proc. 27th Stapp Car Crash Conference, pp. 56–65. Paper No. 831613.
W ISMANS , J., VAN O ORSCHOT , H., W OLTRING , H.J. (1986). Omni-directional human head–neck response.
W ISMANS , J.S.H.M., et al. (1994). Injury Biomechanics (Eindhoven University of Technology), p. 49.
W ITTEK , A., H AUG , E., K AJZER , J. (1999). Hill-type muscle model for analysis of muscle tension on the
human body response in a car collision using an explicit finite element code. JSME Internat. J. Ser. A:
Solid Material Mech. 43, 8–18.
W ITTEK , A., K AJZER , J. (1995). A review and analysis of mathematical models of muscle for application
in the modeling of musculoskeletal system response to dynamic load. In: Högfors, C., Andreasson, G.
(eds.), Proceedings of the 9th Biomechanics Seminar (Center for Biomechanics Chalmers University of
Technology and Göteborg University, Göteborg), pp. 192–216.
W ITTEK , A., K AJZER , J. (1997). Modeling of muscle influence on the kinematics of the head–neck complex
in impacts. Mem. School Engrg. Nagoya University 49, 155–205.
W ITTEK , A., K AJZER , J., H AUG , E. (1999). Finite element modeling of the muscle effects on kinematic
responses of the head–neck complex in frontal impact at high speed. Japan Soc. Automotive Engrg. Rev.
W ITTEK , A., O NO , K., K AJZER , J. (1999). Finite element modeling of kinematics and dynamic response of
cervical spine in low-speed rear-end collisions: mechanical aspects of facet joint injury. J. Crash Preven-
tion Injury Control.
W ITTEK , A., O NO , K., K AJZER , J., Ö RTENGREN , R., I NAMI , S. (2001). Analysis and comparison of reflex
times and electro-myograms of cervical muscles under impact loading obtained using surface and fine-wire
electrodes. IEEE Trans. Biomedical Engrg. 48 (2), 143–153.
W OO , S.L.Y., J OHNSON , G.A., S MITH , B.A. (1993). Mathematical modelling of ligaments and tendons.
J. Biomech. Engrg. 115, 468–473.
W OO , S.L.Y., P ETERSON , R.H., O HLAND , K.J., et al. (1990). The effects of strain rate on the properties of
the medial collateral ligament. Skeletally Immature and Mature Rabbits: A Biomechanical and Histologi-
cal Study. J. Orthop. Res. 8, 712–721.
W YKOWSKI , E., S INNHUBER , R., A PPEL , H. (1998). Finite element model of human lower extremities in a
frontal impact. In: IRCOBI Conference.
YAMADA , H. (1970). In: Evans, F.G. (ed.), Strength of Biological Materials (Williams & Wilkins, Baltimore,
MD).
YANG, J. (1998). Bibliographic study. Report 3CHA/980529/T1/DB, Chalmers University of Technology,
SE-41296 Göteborg, Sweden.
YANG, J., L ÖVSUND, P. (1997). Development and validation of a human body mathematical model for simu-
lation of car-pedestrian collisions, in: IRCOBI Conference, Hannover.
YANG , K.H., Z HU , F., L UAN , F., Z HAO , L., B EGEMAN , P. (1998). Development of a finite element model
of the human neck. In: Proc. 42nd Stapp Car Crash Conference, pp. 1–11. Paper No. 983157.
Y OGANANDAN , N., H AFFNER , M., M ALMAN , D.J., N ICHOLS , H., P INTAR , F.A., J ENTZEN , J., W EIN -
SHEL , S.S., L ARSON , S.J., S ANCES J R ., A. (1989). Epidemiology and injury biomechanics of motor
vehicle related trauma to the human spine. In: Proc. 33rd Stapp Car Crash Conference, pp. 22–242. Paper
No. 892438.
Y OGANANDAN , N., P INTAR , F.A. (1998). Biomechanics of human thoracic ribs. J. Biomech. Engrg. 120,
100–104.
Y OGANANDAN , N., S ANCES J R ., A., P INTAR , F. (1989b). Biomechanical evaluation of the axial compres-
sive responses of the human cadaveric and manikin necks. J. Biomech. Engrg. 111, 250–255.
Y OGANANDAN , N., S RIRANGAM , K., P INTAR , F.A. (2001). Biomechanics of the cervical spine Part 2.
Cervical spine soft tissue responses and biomechanical modeling. Clinical Biomech. 16, 1–27.
Y OO , W.-H., C HOI , H.-Y. (1999). Finite element modeling of human lower extremity. Hongik J. Sci. Tech-
nol. 3, 187–212.
Z AJAC , F.E. (1989). Muscle and tendon: Properties, models, scaling and applications to biomechanics and
motor control. Critical Rev. Biomedical Engrg. 17, 359–411.
452 E. Haug et al.
Z EIDLER , F., S TÜRTZ , G., B URG , H., R AU , H. (1981). Injury mechanisms in head-on collisions involving
glance-off. In: Proc. 25th Stapp Car Crash Conference, pp. 825–860. Paper No. 811025.
Z INK, L. (1997). Simulation von Unfällen mit Airbagauslösung für den Out-of-Position Fall. Diploma Thesis
prepared at TRW Germany, University of Stuttgart, IVK.
Z HANG , L., BAE , J., H ARDY , W.N., M ONSON , K.L., M ANLEY , G.T., G OLDSMITH , W., YANG , K.H.,
K ING , A. (2002). Computational study of the contribution of the vasculature on the dynamic response of
the brain. Stapp Car Crash J. 46, 145–165. Paper No. 2002-22-0008.
Z HANG , L., YANG , K.H., D WARAMPUDI , R., O MORI , K., L I , T., C HANG , K., H ARDY , W.N., K HALIL ,
T.B., K ING , A.I. (2001). Recent advances in brain injury research: A new human head model development
and validation. In: Proc. 45th Stapp Car Crash Conference, pp. 369–394. Paper No. 2001-22-0017.
Z HOU , C., K HALIL , T.B., K ING , A.I. (1996). Viscoelastic response of the human brain to sagittal and lateral
rotational acceleration by finite element analysis. In: International IRCOBI Conference on the Biomechan-
ics of Impact, Dublin, Ireland.
Z HOU , Q., ROUHANA , S.W., M ELVIN , J.W. (1996). Age effects on thoracic injury tolerance. In: Proc. 40th
Z UURBIER , C.J., E VERARD , P., VAN D ER W EES , A.J., H UIJING , P.A. (1994). Length-force characteristics
of the aponeurosis in the passive and active muscle condition and in the isolated condition. J. Bio-
mech. 27 (4), 445–453.
Soft Tissue Modeling for Surgery
Simulation
Hervé Delingette, Nicholas Ayache
INRIA Sophia–Antipolis, 2004, route des Lucioles,
BP 93, 06902 Sophia–Antipolis, France
E-mail addresses: Herve.Delingette@sophia.inria.fr (H. Delingette),
Nicholas.Ayache@inria.fr (N. Ayache)
Foreword
In this chapter, we address a specific issue belonging to the field of biomechanics –

modeling living tissue deformation with real-time constraints. This issue was raised by
the emergence, in the middle of the 1990s, of a very specific application – the simula-
tion of surgical procedures. This new concept of surgery simulation was in large part
advocated by the American Department of Defense (S ATAVA [1994]), for which this
concept was a key part of their vision of the future of emergency medicine.
Since then, the concept of having surgeons being trained on simulators (just like pilots
on flight simulators) has been refined. First, the development of commercial simulators
has proved that there was a demand for products that help to optimize the learning
curve of surgeons.1 Second, the emergence of medical robotics and more precisely of
minimally invasive surgery robots, has reinforced the need for simulating surgical pro-
cedures, since these robots require a very specific hand–eye coordination. Third, there is
a large consensus among the medical community that current simulators are not realistic
enough to provide advanced gesture training. In particular, the modeling of living tissue,
and their ability to deform under the contact of an instrument is one of the important
aspect of simulators that should be improved.
1 This curve represents the number of incidents occurring during the performance of surgery as a function
of time. This curve is generally monotonically decreasing under the effect of training and usually reaches an
asymptotic value after a certain number N of real interventions. The objective of the simulators is to reduce
this number N as much as possible.
453
454 H. Delingette and N. Ayache
In this chapter, we present different algorithms for modeling soft tissue deforma-
tion in the context of surgery simulation. These algorithms make radical simplifications
about tissue material property, tissue visco-elasticity and tissue anatomy. The first sec-
tion of this chapter describes the principles and the components of a surgical simulator.
In particular, we insist on the different constraints of soft tissue models in this applica-
tion, the most important being the real-time computation constraint. In Section 2, we
present the process of building a patient-specific hepatic surgery simulator from a set of
medical images. The different stages of computation leading to the creation of a volu-
metric tetrahedral mesh from a medical image are especially emphasized. In Section 3,
we detail the five main hypotheses that are made in the proposed soft tissue models. Fur-
thermore, we recall the main equations of isotropic and transversally anisotropic linear
elasticity in continuum mechanics. The discretization of these equations are presented
in Section 4 based on finite element modeling. Because we rely on the simple linear
tetrahedron element, we provide closed form expressions of local and global stiffness
matrices. After describing the types of boundary conditions existing in surgery simu-
lation, we derive the static and dynamic equilibrium equations in their matrix form. In
Section 5, a first model of soft tissue is proposed. It is based on the off-line inversion of
the stiffness matrix and can be computed very efficiently as long as no topology change
is required. In such case, in Section 6, a second soft tissue model allows to perform
cutting and tearing but with less efficiency as the previous model. A combination of the
two previous models, called “hybrid model” is also presented in this section. Finally, in
Section 7, we introduce an extension of the second soft tissue model that implements
large displacement elasticity.
1. General issues in surgery simulation
1.1. Surgical simulators
1.1.1. Medical impact of surgical simulators

Surgery simulation aims at reproducing the visual and haptic senses experienced by a
surgeon during a surgical procedure, through the use of computer and robotics systems.
The medical interest of this technology is linked with the development of minimally
invasive techniques especially video-surgery (endoscopy, laparoscopy, . . . ). More pre-
cisely, laparoscopy consists in performing surgery by introducing different surgical in-
struments in the patient abdomen through one centimeter-wide incisions. The surgeon
can see the abdominal anatomy with great clarity by watching a high resolution mon-
itor connected to an endoscope introduced inside the patient abdomen. This technique
bears several advantages over traditional open surgery. On one hand, it decreases the
trauma entailed by the surgical procedure on the patient body. This allows to decrease
the patient stay in hospitals and therefore decreases the cost of health care. On the other
hand, it reduces the morbidity as demonstrated by the Hunter and Sackier study (B ERCI ,
H UNTER and S ACKIER [1994]).
However, if these minimally invasive techniques are clearly beneficial to the patients,
they also bring new constraints on the surgical practice. First, they significantly degrade
Soft tissue modeling for surgery simulation 455
the surgeon access to the patient body. In laparoscopy, for instance, the surgical proce-
dure is made more complex by the limited number of degrees of freedom of each sur-
gical instrument. Indeed, they must go through fixed points where the incisions in the
patient’s abdomen were done. Furthermore, because the surgeon cannot see his hand
on the monitor, this technique requires a specific hand–eye coordination. Therefore, an
important training phase is required before a surgeon acquires the skills necessary to ad-
equately perform minimally invasive surgery (corresponding to a plateau in the learning
curve).
Currently, surgeons are trained to perform minimally invasive surgery by using me-
chanical simulators or living animals. The former method is based on “endotrainers”
representing an abdominal cavity inside which are placed plastic objects representing
human organs. These systems are sufficient for acquiring basic surgical skills but are
not realistic enough to represent fully the complexity of the human anatomy and physi-
ology (respiratory motion, bleeding, . . . ). The latter training method consists in practic-
ing simple or complex surgical procedures on living animals (often pigs for abdominal
surgery). This method has two limitations. First, the similarity between the human and
animal anatomy is limited and therefore certain procedures cannot be precisely simu-
lated with this technique. Also, the evolution of the ethical code in most countries may
forbid the use of animals for this specific training, as it is already the case in several
European and North American countries.
Because of the limitations of current training methods, there is a large interest in
developing video-surgery simulation software for providing efficient and quantitative
gesture training systems (AYACHE and D ELINGETTE [2003]). Indeed, such systems
should bring a greater flexibility by providing scenarios including different types of
pathologies. Furthermore, thanks to the development of medical image reconstruction
algorithms, surgery simulation allows surgeons to verify and optimize the surgical strat-
egy of a procedure on a given patient.
1.1.2. Classification of surgical simulators

S ATAVA [1996] et al. proposed to classify surgical simulators into three categories (see
Fig. 1.1). The first generation simulators are solely based on anatomical information,
in particular on the geometry of the anatomical structures included in the simulator.
In these simulators, the user can virtually navigate inside the human body but has a
limited interaction with the modeled organs. Currently, several first generation surgi-
cal simulators are available including commercial products linked to medical imaging
systems (CT or MRI scanners) that are focusing on virtual endoscopy (colonoscopy,
tracheoscopy, . . . ). In general, they are used as a complementary examination tools
establishing a diagnosis (for instance, when using virtual endoscopy) or as a surgical
planning tool before performing surgery.
In addition to geometrical information, second generation simulators describe the
physical properties of the human body. For instance, the modeling of soft tissue biome-
chanical properties enables the simulation of basic surgical gestures such as cutting or
suturing. Currently, several prototypes of second generation simulators are being devel-
oped including the simulation of cholecystectomy (C OVER , E ZQUERRA and O’B RIEN
[1993], K UHN , K ÜHNAPFEL , K RUMM and N EISIUS [1996]), of arthroscopy of the
F IG . 1.1. The different generations of medical simulators.
knee (G IBSON , S AMOSKY, M OR , F YOCK , G RIMSON , K ANADE , K IKINIS , L AUER

and M C K ENZIE [1997]) and of gynecological surgery (S ZEKELY, BAIJKA and B RECH -
BUHLER [1999]). Section 2 will shortly describe the hepatic surgery simulator being
developed at INRIA.
Third generation of surgical simulators provides an anatomical, physical and physi-
ological description of the human body. There are very few simulators including these
three levels of modeling, essentially because of the difficulty to realistically describe
the coupling between physiology and physics. A good example of an attempt in this
direction is given by the work of K AYE , P RIMIANO and M ETAXAS [1997] who mod-
eled the mechanical cardiopulmonary interactions. Another important example is the
study of the contraction of the right and left ventricles of the heart under the propaga-
tion of the action potential which is being carried out by the group of Prof. McCulloch
(this work is published in this book) but also by the INRIA ICEMA group (S ERME -
SANT, C OUDIÈRE , D ELINGETTE and AYACHE [2002], S ERMESANT, FARIS , E VANS ,
M C V EIGH , C OUDIÈRE , D ELINGETTE and AYACHE [2003]). Finally, it should be noted
that a comprehensive effort for creating computational physiological models has been
recently launched in the international Physiome Project (BASSINGTHWAIGHTE [2000]).
1.2. Simulator architecture
In this section, we present the basic components of simulators for surgical gesture train-
ing and especially in the context of minimally invasive therapy. For the acquisition of
basic skills, it is necessary to simulate the behavior of “living” tissues and therefore to
develop a second generation surgical simulator. However, it raises important technical
and scientific issues. The different components of these simulators are summarized in
Fig. 1.2.
F IG . 1.2. The different components of a second generation surgery simulator.
The input devices in such simulators usually consist of one or several mechanical
systems that drive the motion of virtual surgical tools or of virtual endoscopes. In fact,
as input devices they do not need to be motorized and they are usually equipped with
simple optical encoders or position trackers. A keyboard and electronic mouse are also
useful to modify the scenario of the simulation.
The core of a simulator consists of several modules. For instance, a first module
provides the enabling tools for the creation of geometric models from medical images
(see Section 1.2.1). Another module, detailed in Section 1.3, computes the deformation
of soft tissues under the action of virtual instruments. These interactions between vir-
tual instruments and virtual organs, performed in a separate module, mainly consists
in detecting collisions followed with modeling contact forces and displacements (see
Section 1.2.2).
Finally, a surgical simulator must provide an advanced user interface that includes
visual and force feedback (respectively presented in Sections 1.2.3 and 1.2.4). Last but
not least, it is necessary to rely on advanced software engineering methodology to make
these different modules communicate within the same framework: some of these imple-
mentation issues are introduced in Section 1.2.5.
1.2.1. Geometric modeling

In general, the extraction of tridimensional geometric models of anatomical structures
is based on medical imagery: CT scanner images, MRI images, cryogenic images, 3D
ultrasound images, . . . . Because medical image resolution and contrast have greatly
improved over the past few years, the tridimensional reconstruction of certain struc-
tures have become possible by using computerized tools. For instance, the availability
in 1995 of the “Visible Human” dataset provided by the National Library of Medicine
has allowed the creation of a complete geometric human model (ACKERMAN [1998]).
However, the automatic delineation of structures from medical images is still consid-
ered an unsolved problem. Therefore, a lot of human interaction is usually required for
reconstructing the human anatomy. D UNCAN and AYACHE [2000], AYACHE [2003],
provide a survey on the past and current research effort in medical image analysis.
1.2.2. Interaction with a virtual instrument

A key component of a surgery simulation software is the user interface. The hardware
interface that drives the virtual instrument essentially consists in one or several force-
feedback devices having the same degrees of freedom and appearance as the actual
surgical instruments used in minimally invasive therapy (see Fig. 1.3). In general, these
systems are force-controlled, sending the instrument’s position to the simulation soft-
ware and receiving reaction force vectors.
F IG . 1.3. A force feedback system suited for surgery simulation.

Once the position of the virtual instrument is known, it is necessary to detect pos-
sible collisions with other instruments or surrounding anatomical structures. In this
case, it is particularly difficult to obtain a computationally efficient collision detec-
tion algorithm because the geometry of objects may change at each iteration. There-
fore, algorithms based on pre-computed data structures (such as the approach pro-
posed in G OTTSCHALK , L IN and M ANOCHA [1996]) are not appropriate. L OMBARDO ,
C ANI and N EYRET [1999] proposed an original collision detection method based on
the OpenGL graphics library which is especially well-suited for elongated instruments
shaped like those used in laparoscopic surgery. Although this technique cannot be
used for the detection of self-collisions, several algorithms have been proposed re-
cently (T ESCHNER , H EIDELBERGER , M ULLER , P OMERANETS and G ROSS [2003],
K NOTT and PAI [2003]) to tackle this complex task.
When a collision is detected, a set of geometrical or physical constraints are applied
on soft tissue models. However, modeling the physics of contacts can lead to complex
algorithms and therefore purely geometric approaches are often preferred.
1.2.3. Visual feedback

A surgery simulator must provide a realistic visual representation of the surgical pro-
cedure. Visual feedback is especially important in video-surgery because it helps the
surgeon to acquire a tridimensional perception of his environment. In particular, the ef-
fects of shading, shadows and textures are important clues that must be reproduced in a
simulator.
The quality of visual feedback is directly related to the availability and performance
of graphics accelerators. In the past few years, the market of graphics cards has evolved
in three directions: improved price-performance ratio, increased geometric transforma-
tion and rasterization performance and the emergence of programmable pixel rendering.
Combined with the development of more efficient computer graphics algorithms, we can
foresee that realistic visual feedback for surgery simulation could be achieved in the next
few years if this additional graphics rendering is focused on the three-dimensional clues
used by surgeons to understand the surgical field.
1.2.4. Haptic feedback

Haptic display serves at least two purposes in a surgical simulator: kinesthetic and cog-
nitive. First, it provides the sensation of movement to the user and therefore it signif-
icantly enhances surgical performance. Second, it helps to distinguish between tissues
by testing their mechanical properties.
However, the addition of a haptic display in a simulation system increases by a
large factor its complexity and the required computational power (M ARK , R ANDOLPH ,
F INCH , VAN V ERTH and TAYLOR II [1996]): it leads to an increase by a factor 10
of the required bandwidth, synchronisation between visual and haptic displays, force
computation, . . . . Only a few papers have assessed the importance of haptic feedback
in surgery (M ARCUS [1996]). In general, it is accepted that the combination of visual
and haptic displays is optimal for surgery training or pre-planning.
In video-surgery, the surgical instruments slide inside a trocard and are constrained
to enter the abdomen through a fixed point. This entails substantial friction, specifically
for laparoscopy where airtightness must be enforced. The friction of the instruments
inside trocards perturbes the forces sensed by the end-user. Despite those perturbations,
it appears that it is still necessary to provide force-feedback for realistic user immersion.
1.2.5. Implementation of a simulator

Most of the difficulties encountered when implementing a surgical simulator originate
from the trade-off that must be found between real-time interaction and the necessary
surgical realism of a simulator.
The first constraint indicates that a minimum bandwidth between the computer and
the interface devices must be available in order to provide a satisfactory visual and
haptic feedback. If this bandwidth is too small, the user cannot properly interact with
the simulator and it becomes useless for surgery gesture training. However, the “real-
time” constraint can be interpreted in different ways. Most of the time, it implies that the
mean update rate is high enough to allow a suitable interaction. However, it is possible
that during the simulation, some events (such as the collision with a new structure)
may increase the computational load of the simulation engine. This may result in a
lack of synchronicity between the user gesture and the feedback the user gets from
the simulator. When the computation time is too irregular, the user may even not be
able to use the simulator properly. In order to guarantee a good user interaction, it is
necessary to use a dedicated “real-time” software that supervises all tasks executed on
the simulator.
The second constraint is related to the targeted application of a simulator: training
surgeons to new gestures or procedures. To reach this goal, the user must “believe” that
the simulator environment corresponds to a real procedure. The level of realism of a
simulator is therefore very dependent on the type of surgical procedures and is also
connected with physio-psychological parameters. In any case, increasing the realism of
a simulator requires an increase of computational time which is contradictory with the
constraint of real-time interaction.
The main difficulty in implementing a simulator is to optimize its credibility, given an
amount of graphics and computational resources. Therefore, an analysis of the training
scenario should be performed to find the most important elements that contribute to the
realism of the simulation.
1.3. Constraints of soft tissue models
In the scope of a surgical simulator, it is not possible to model the biomechanical com-
plexity of living soft tissue. Instead, authors have resorted to simplified models to de-
crease the implementation complexity and to optimize computational efficiency. A sur-
vey on soft tissue modeling can be found in D ELINGETTE [1998].
Before presenting the main features of our approach (available in Section 3.1), we list
three constraints that should be taken into account when specifying a soft tissue model
for surgery simulation.
1.3.1. Visualization constraints

To obtain high quality visual rendering, two techniques are traditionally used: sur-
face and volume rendering. A comparison between these two rendering techniques for
surgery simulation is described in S OFERMAN , B LYTHE and J OHN [1998]. Surface ren-
dering is by far the most commonly used technique, and uses basic polygonal elements
(triangles, quads, . . . ) to achieve the rendering of an entire scene. A rule of thumb in
surface rendering states that the quality of rendering is proportional to the number of
polygonal elements. Unfortunately, the screen refresh rate of a graphics display is in-
versely proportional to the number of elements.
Therefore, an important concern arises when specifying a soft tissue model: is it com-
patible with high quality visual rendering? For some models, it is clearly not the case.
For instance, the chain–mail algorithm (G IBSON , S AMOSKY, M OR , F YOCK , G RIM -
SON , K ANADE , K IKINIS , L AUER and M C K ENZIE [1997]) represents soft tissue with
the help of cubic lattices that are allowed to move slightly with respect to their neigh-
bors. For this representation, as well as for particle-based representations (F RANCE ,
A NGELIDIS , M ESEURE , C ANI , L ENOIR , FAURE and C HAILLOU [2002], D ESBRUN
and G ASCUEL [1995]) and multigrid representations (D EBUNNE , D ESBRUN , C ANI and
BARR [2001]), authors use a two-layer strategy: a volumetric soft tissue model is com-
bined with a surface model dedicated to visual rendering. These two models are often
coupled with a linear relationship based on barycentric coordinates: once the shape of
a soft tissue model is modified, the surface model is updated in an efficient manner.
Similarly, the collision detection is performed on the surface model, but contact forces
and displacements are imposed on the volumetric model. However, this approach has
two limitations. First, the modeling of contact between a virtual tool and a soft tissue
model is usually not satisfactory because the mapping between surface and volumetric
model is complex (though mapping from volumetric to surface models is often trivial).
Second, this approach makes the modeling of tissue cutting very complex where the
surface and volumetric topology is altered.
For soft tissue models based on tetrahedral or hexahedral meshes, the problem of
high quality visual rendering is posed in a different manner since the shell of these
meshes (made of triangular or quadrangular elements) can be used directly for ren-
dering. However, in general, coarse volumetric meshes are used in order to achieve
real-time performances (see next section). Therefore, it is often required to compen-
sate the poor geometrical quality by using specific computer graphics algorithms such
as subdivision surfaces (Z ORIN , S CHROEDER and S WELDENS [1996]), using avatars
(D ECORET, S CHAUFLER , S ILLION and D ORSEY [1999]) or by replacing elements with
texture (S ILLION , D RETTAKIS and B ODELET [1997]). In the case of the hepatic surgery
simulator, we have used the PN triangles algorithms (V LACHOS , P ETERS , B OYD and
M ITCHELL [2001]) in order to provide a smooth visual rendering of the liver. The idea
behind PN triangles is to subdivide each triangle and its normal vector into subtriangles
in order to produce a smoother looking surface (see Fig. 1.4 for an example).
1.3.2. Real time deformation constraints

A surgical simulator is an example of a virtual reality system. To succeed in training
surgeons, a simulator must provide an advanced user interface that leads to the immer-
sion of surgeons into the virtual surgical field. To reach this level of interaction, three
basic rules must be formulated:
F IG . 1.4. Display of a liver being resected: (a) display of the triangles corresponding to the shell of the liver
tetrahedral mesh; (b) surface rendering based on Gouraud shading without PN triangles; (c) surface rendering
based on PN triangles with two levels of subdivisions.
Rule 1. Minimum bandwidth for visual and haptic feedback. An acceptable bandwidth
for visual display is in the range of 20–60 Hz while the acceptable bandwidth for
haptic display is on the range of 300–1000 Hz (300 Hz is the free-hand gesture fre-
quency). In fact, this notion of minimal bandwidth depends on the nature of the scene
to be displayed: for objects moving slowly on the screen, an update rate of 20 Hz is
sufficient. Similarly, a frequency of 300 Hz may be enough to render the contact with
very soft objects.
Rule 2. Low latency. Latency measures the time between measurements performed by
the sensor (for instance, the position of the surgical instrument) and action (visual or
haptic display). Latency is critical for user immersion. The hardware configuration of
the system can greatly influence latency since communication between elements may
be responsible for additional delays. In Fig. 1.5, the architecture of the simulation
system used at INRIA (C OTIN , D ELINGETTE , C LEMENT, TASSETTI , M ARESCAUX
and AYACHE [1996]) in 1996 is presented. It is composed of one haptic display, a per-
sonal computer and a graphics workstation. There are several causes contributing to
latency: communication delays between the haptic display and the PC, communica-
tion between the PC and the graphics workstation, the delay caused by the graphics
display, the computation time for collision detection, force feedback and deforma-
tion. Since some of the communication links between elements are asynchronous,
the total latency is not the sum of those delays but it is important to reduce them
to their minimum values. The latency depends greatly on hardware, specifically on
computation and graphics performance.
Rule 3. Realistic motion of soft tissue. It is important that the dynamic behavior of a
deformable tissue is correctly simulated. To assess the visco-elastic behavior of a
material, one can measure the speed at which an object returns to its rest position
after it is perturbed. Soft tissue undergoes a damped motion whereas stiff objects
react almost instantaneously to any perturbation. At the limit, very stiff objects can be
considered to have a quasi-static motion, implying that static equilibrium is reached
at each time-step (see Section 5 for a discussion about quasi-static motion).
In terms of soft tissue modeling, two parameters are important for real-time deformation
constraints. The first parameter is the update frequency fu which controls the rate at
which the shape of a soft tissue model is modified. If we write Xt as the position of
F IG . 1.5. Architecture of a simulator composed of a personal computer driving an haptic device and a graph-
ics workstation.
the tissue model at iteration t, the computation time Tc = 1/fu is the time needed to
compute the new position Xt +1 . The second parameter is the relaxation time Trelaxation
which is the time needed for a material to return to its rest position once it has been
perturbed.
To reach the required bandwidth for haptic and visual rendering (Rule 1) it is nec-
essary that the computation time Tc is bounded by a constant Tinteraction that depends
on the architecture of the simulator. For instance, in Fig. 1.6 we display three different
software architectures for handling soft tissue deformation, visual and haptic feedback.
In a first architecture (Fig. 1.6(a)), all three tasks are performed sequentially, one
after the other. The advantage of this approach lies in its simplicity of implementation.
However, it has two drawbacks. The main problem is that the computation time Tc must
be short enough to follow the minimum frequency for haptic feedback: 300 to 1000 Hz.
This implies that Tinteraction ≈ 2 ms which is a very high requirement for a soft tissue
model of reasonable size. In fact, to the best of our knowledge, only methods based on
pre-computation of the static solution such as the one proposed in Section 5 can comply
with this constraint. The second problem with this approach is that a delay in any of the
three tasks perturbs the other tasks. For instance, when the user performs tissue cutting,
an additional task is needed to update the mesh topology which translates into a delay
in the visual and haptic rendering.
The second architecture shown in Fig. 1.6(b) is the most commonly used in today’s
surgical simulators: the haptic rendering is performed in a different process or different
thread than the visual rendering and soft tissue modeling tasks. Its purpose is to sharply
decrease the real-time constraint on the soft tissue computation from haptic frequency
(≈ 500 Hz, Tinteraction ≈ 2 ms) to visual frequency (≈ 25 Hz, Tinteraction ≈ 40 ms). In
F IG . 1.6. Different software architecture for handling visual rendering, haptic rendering and soft tissue mod-
eling.
order to keep a satisfactory force feedback, a separate thread or process, running at

haptic frequency, computes the force intensity for the haptic device based on a sim-
plified local model. This local model, that may consist of a sphere (S ERRANO and
L AUGIER [2001]) or a plane (F OREST, D ELINGETTE and AYACHE [2002a]) is updated
by the soft tissue modeling loop while the position of the virtual surgical tool, neces-
sary to compute its contact with soft tissue, is updated by the haptic rendering process
and sent to the process. This asynchronous communication between haptic and visual
rendering gives satisfactory results when some temporal smoothing is performed dur-
ing the computation of force intensity. The main drawback of this approach is that it
increases software complexity compared to the previous architecture. However, since
only little information must be shared between the two processes, it has been adopted
in several simulators, including the current version of the INRIA hepatic surgery simu-
lator.
In the third architecture described in Fig. 1.6(c), the visual and haptic rendering
tasks are performed in separate processes or threads in order to remove the latency
caused by graphics hardware. Furthermore, this architecture makes the computation of
soft tissue deformation more efficient (decrease of Tc ) when compared to previous so-
lutions. However, it has little effect on the maximum computation time per iteration
Tinteraction ≈ 40 ms since the geometric model still requires to be updated at 25 Hz for
visual rendering. This approach is more difficult to implement because the amount of
information to transmit to the visual rendering task is quite large. Furthermore, a change
in mesh topology during simulation requires to modify the data structure of the com-
puter graphics algorithm. An example of such architecture is provided in B IELSER and
G ROSS [2000].
To summarize, we can state that a soft tissue model in a surgical simulator must
essentially meet two constraints: computation time Tc per iteration less than a constant
Tinteraction, and relaxation time Trelaxation defined by the visco-elastic behavior of the
material.
1.3.3. Tissue cutting and suturing

The ability to cut and suture tissue is of primary importance for designing a surgery
simulation system. The impact of such operations in terms of tissue modeling is consid-
erable since it implies changing tissue topology over time. The cost of such a topological
change depends largely on the chosen geometric representation but also on the numer-
ical method that is adopted to compute tissue deformation (see discussion in previous
section).
In addition, the tissue behavior must be adapted at locations where cutting or su-
turing occurs. Little is known about the stress/strain relationship occurring during and
after cutting. The basic assumption that is made is that the physical properties of tis-
sue are only modified locally. However, in practice, cutting can modify the boundary
conditions significantly between tissue and the surrounding organs which implies con-
siderable change with respect to their ability to deform.
Finally, when cutting volumetric or surface models, it is very likely that the new
geometric and physical representation of tissue leads to self-intersections. The detection
of self-intersections is very computationally expensive and, therefore repulsive force
between neighboring vertices are sometimes added to prevent self-intersections.
1.4. Computational methods for soft tissue modeling
Several computational methods can be employed for modeling the deformation of soft
tissue. We simplify the taxonomy of these methods by proposing the three classes of
algorithms most commonly used:
• Direct methods. This category contains all methods that solve the static or dy-
namic equilibrium equation at each iteration (quasi-static motion). To reach such
performance, some kind of pre-computation is performed. The algorithm presented
in Section 5 is a direct method as well as the algorithm described in D EBUNNE ,
D ESBRUN , C ANI and BARR [2001], R ADETZKY [1998].
• Explicit iterative methods. With iterative methods, the deformation is computed
as the limit (in finite time) of a converging series that have been initialized.
The closer the initial value is from the solution the faster the convergence. Iter-
ative methods can be implemented based on implicit or explicit schemes. With
TABLE 1.1
Comparison between the three soft tissue models: direct methods (pre-computed quasi-static model), explicit
iterative schemes (tensor–mass and spring–mass models) and implicit iterative schemes (Houbolt or Newmark
methods)
Direct methods Explicit iterative scheme Implicit iterative scheme
Computation time low low high

Relaxation time low high low
Cutting simulation very difficult possible difficult
explicit schemes, the next position of the tissue model Xt +1 is obtained from
the application of internal forces estimated at iteration t. These methods encom-
pass the most common algorithms found in the literature for modeling soft tis-
sue deformation, including spring–mass model (K UHNAPFEL , A KMAK and M AA
[2000]), tensor–mass models (C OTIN , D ELINGETTE and AYACHE [2000]) (pre-
sented in Section 6), the “chain–mail” algorithm (G IBSON , S AMOSKY, M OR , F Y-
OCK , G RIMSON , K ANADE , K IKINIS , L AUER and M C K ENZIE [1997]) and others
(B RO -N IELSEN [1998]).
• Semi-implicit iterative methods. With implicit or semi-implicit schemes, the next
position of the tissue model Xt +1 is obtained from the application of internal forces
estimated at iteration t + 1. Therefore, a linear system of equations needs to be
solved entirely or partially (BARAFF and W ITKIN [1998]).
In Table 1.1, we present the general features of these three types of numerical methods
with respect to the constraints enumerated in Section 1.3. More precisely, the time of
computation, the relaxation time (inversely proportional to the speed of convergence
towards the rest position) and the ability to support any change of mesh topology during
the simulation of cutting or suturing is estimated qualitatively for each method.
Direct methods can support high frequency update fu and may have a low relaxation
time to model stiff material, but they cannot simulate tissue cutting since they rely on
the precomputation of some parameters.
On the other hand, explicit iterative methods are well-suited for the simulation of
cutting, but these method often suffer from a high relaxation time, which makes their
dynamic behavior somewhat unrealistic (jelly-like behavior). This high relaxation time
originates from a lack of synchronicity, where the time step t used in the discretiza-
tion of the explicit scheme, is much smaller than the computation time Tc . To obtain
satisfactory results, it is often required to use a mesh with a small number of nodes
(typically less than 1000 vertices on a standard PC).
Finally, with implicit iterative methods, the time step t can be increased by an
order of magnitude compared to the explicit case. This allows to obtain much better
dynamical behavior, but, on the other hand, they suffer from a higher computation time
than explicit methods since a (sparse) linear system of equations needs to be solved at
each iteration. Again, to achieve real-time performance, these methods are limited to
meshes with a small number of vertices.
2. The INRIA hepatic surgery simulator
2.1. Objectives
In the sequel we use the hepatic surgery simulator developed at INRIA in the Epidaure
project2 as a case study to illustrate the different algorithms and the practical issues
involved when building soft tissue models.
The INRIA hepatic surgery simulator was initiated in 1995 as a part of the European
project MASTER in collaboration with the IRCAD research center3 which hosts the
European Institute of Tele-Surgery (EITS). The motivations that led us to propose the
development of an hepatic surgery simulator were twofold.
First, hepatic pathologies are among the major causes of death worldwide. For in-
stance, hepatocellular carcinoma (HCC) is a primary liver cell cancer and it accounts
for most of cancer tumors. It causes the death of 1 250 000 people mainly in Asia and
Africa. Furthermore, hepatic metastases (secondary tumorous cells) are mainly caused
by colorectal cancers (in 30 to 50% of cases) and patients have little chance to survive
hepatic carcinoma without any therapy (0 to 3% of survival for a 5 year period with an
average survival period of 10 months).
The second motivation is related to the nature of hepatic resection surgery. Indeed,
this surgical procedure involves many generic surgical gestures (large displacement mo-
tion, grasping, cutting, suturing) that can be useful in the simulation of different proce-
dures. Also, because of the presence of hepatic parenchyma, the tissue models must be
of volumetric nature which departs significantly from previously developed simulators
simulating hollow organs like the gall bladder. Finally, tissue being a soft material al-
lows to employ low-end force feedback systems for simulating contact forces between
surgical tools and hepatic tissue.
This work has greatly benefited from the INRIA incentive action AISIM4 which
gathered different INRIA teams working in the fields of medical image analysis (Epi-
daure), robotics (Sharp) (B OUX DE C ASSON and L AUGIER [1999]), computer graphics
(Imagis) (D EBUNNE , D ESBRUN , C ANI and BARR [2001]) and numerical analysis (Si-
nus, Macs) (V IDRASCU , D ELINGETTE and AYACHE [2001]).
2.2. Liver anatomy
The liver is the largest gland (average length of about 28 cm, average height of about
16 cm and average greatest thickness of about 8 cm) in the human body. It has numer-
ous physiological functions: to filter, metabolize, recycle, detoxify, produce, store and
destroy. It is located in the right hypochondriac and epigastric regions (see Fig. 2.1).
The liver has a fibrous coat, the so-called Glisson’s capsule. Its rheological behavior is
quite different from the glandular parenchyma. Five vessel types run through the liver
parenchyma: biliary and lymphatic ducts on one hand, blood vessels (internal portal
2 Description of the Epidaure project is provided at http://www.inria.fr/epidaure/.
3 Institut de Recherche Centre le Cancer de l’Appareil Digestif, 1, place de l’Hôpital, 67091 Strasbourg
cedex, France, http://www.ircad.com/, funded by Prof. J. Marescaux.
4 http://www-sop.inria.fr/epidaure/AISIM/ .
F IG . 2.1. Description of the liver anatomy with its neighboring structures (source Children’s Liver Disease
Foundation).
supply, hepatic arterial tree and collecting venous network) on the other hand. The por-
tal vein, which conveys blood from the digestive tract to be detoxified and metabolized,
is deep to the proper hepatic artery and common bile duct. This hepatic triad runs to the
liver; it enters the liver via the hilum. This region is thus supposed to be wholly stable.
2.3. Creation of an anatomical model of the liver
In order to produce a model of the liver with anatomical details, the Visible Human
dataset (ACKERMAN [1998]) provided by the National Library of Medicine was used.
This dataset consists of axial MRI images of the head and neck and longitudinal sections
of the rest of the body. The CT data consists of axial scans of the entire body taken at
1 mm intervals. The axial anatomical images are scanned pictures of cryogenic slices
of the body. They are 24-bit color images whose size is 2048 × 1216 pixels. These
anatomical slices are also at 1 mm interval and are registered with the CT axial images.
There are 1878 cross-sections for each modality.
To extract the shape of the liver from this dataset, we used the anatomical slices (cf.
Fig. 2.2), which give a better contrast between the liver and the surrounding organs. The
dataset corresponding to the liver is reduced to about 180 slices. After contrast enhance-
ment, we apply an edge detection algorithm to extract the contours of the image, and
then using a simple thresholding technique, we retain only those with higher-strength
contours are considered for further processing. Next, we use semi-automatic deformable
contour (K ASS , W ITKIN and T ERZOPOULOS [1988], D ELINGETTE and M ONTAGNAT
[2001]) to extract a smooth two-dimensional boundary of each liver slice. These con-
tours are finally transformed into a set of two-dimensional binary images (cf. Fig. 2.2).
The slices generated are then stacked to form a tridimensional binary image (M ONTAG -
NAT and D ELINGETTE [1998]) (cf. Fig. 2.3).
In order to capture the shape of the external surface of the liver, one could use a sub-
voxel triangulation provided by the marching-cubes algorithm (L ORENSEN and C LINE
F IG . 2.2. Slice-by-slice segmentation of the liver. The initial data (left) is a high resolution photography of an
anatomical slice of the abdomen. The binary image (right) corresponds to the segmented liver cross-section.
F IG . 2.3. After segmentation, the binary images are stacked (left) to give a 3D binary image. We see the
step-effect on the shape of the liver (right) when extracted using the marching-cubes algorithm.
[1987]), however the number of triangles generated is too large for further process-
ing. Moreover, a smoothing of the surface is necessary to avoid staircase effects (see
Fig. 2.3). A possible solution consists in decimating an iso-surface model by using a
mesh simplification tool (S CHROEDER , Z ARGE and L ORENSEN [1992]). However, for
more flexibility, in both the segmentation and simplification processes, liver reconstruc-
tion was performed using simplex meshes.
Simplex meshes are an original representation of tridimensional objects developed
by D ELINGETTE [1999]. A simplex mesh is a deformable discrete surface mesh that
is well-suited for generating geometric models from volumetric data. A simplex mesh
can be deformed under the action of regularizing and external forces. Additional prop-
erties like a constant connectivity between vertices and a duality with triangulations
have been defined. Moreover, simplex meshes are adaptive, for example, by concentrat-
ing vertices in areas of high curvature (thereby achieving an optimal shape description
for a given number of vertices). The mesh may be refined or decimated depending on
the distance of the vertices from the dataset. The decimation can also be interactively
controlled. Fig. 2.4 shows the effect of the mesh adaptation and where the vertices are
nicely concentrated at highly high curvature locations of the liver.
By integrating simplex meshes in the segmentation process, we have obtained
smoothed triangulated surfaces, very close to an iso-surface extraction, but with fewer
faces to represent the shape of the organs. In our example, the liver model has been cre-
ated by fitting a simplex mesh to the tridimensional binary image previously described.
F IG . 2.4. Different representations of the geometric liver model. The simplex mesh (M ONTAGNAT and
D ELINGETTE [1998]) fitting the data (top left) with a concentration of vertices in areas of high curvature,
the triangulated dual surface (top right) and a texture-mapped model with anatomical details (gall bladder and
ducts) from an endoscopic viewpoint (bottom).
F IG . 2.5. (a) Original CT-scan images of the liver; (b) reconstructed liver model; (c) outline of the liver sur-
face model in a CT-scan image; (d) segmentation of the portal vein (S OLER , D ELINGETTE , M ALANDAIN ,
M ONTAGNAT, AYACHE , C LÉMENT, K OEHL , D OURTHE , M UTTER and M ARESCAUX [2000]); (e) recon-
struction of the eight anatomical segments (Couinaud segmentation).
Thanks to the adaptation and decimation properties of the simplex meshes, this model
is composed of only 14 000 triangles, whereas the marching-cubes algorithm produced
94 000 triangles (cf. Figs. 2.3 and 2.4).
Although this approach is very useful for building a “generic” liver model, it is es-
sential to integrate “patient-based” models in the simulator. In the framework of this
research project, M ONTAGNAT and D ELINGETTE [1998] have developed a method
for extracting liver models from CT scan images. The principle of this algorithm is
to deform a generic simplex mesh (for instance, the one extracted from the Visible
Human dataset) such that its surface coincides with the liver boundary in the image.
The work of S OLER , M ALANDAIN and D ELINGETTE [1998], S OLER , D ELINGETTE ,
M ALANDAIN , M ONTAGNAT, AYACHE , C LÉMENT, KOEHL , D OURTHE , M UTTER and
M ARESCAUX [2000] has extended this work by additionally extracting the main bifur-
cations of the portal and hepatic veins but also the hepatic lesions and gall-bladder (see
Fig. 2.5).
2.4. Liver boundary conditions
In the scope of the AISIM project, a reference liver model was created by V IDRASCU ,
D ELINGETTE and AYACHE [2001]. They define the liver environment (V IDRASCU ,
D ELINGETTE and AYACHE [2001]) in order to set up the boundary conditions asso-
ciated to computational models. The right liver extremity is thick and rounded while
the left one is thin and flattened. Both extremities are not submitted to specific loads.
The anterior border is thin, sharp and free. The posterior border is connected to the di-
aphragm by the coronary ligament. The upper surface, covered by the peritoneum, is
divided into 2 parts by the suspensory ligament. However, this ligament does not affect
the biomechanical behavior of the liver. The lower surface is connected with the gall-
blader (GB) within the GB fossa, the stomach, the duodenum, the right kidney and the
right part of the transverse colon. These organs are in contact with the liver surface, but
they do not interact strongly with the liver; they cannot be considered as being support-
ing organs. The inferior vena cava (IVC) travels along the posterior surface, very often
in a groove. The connection implies another strong fitting condition (clamp).
2.5. Material characteristics
The literature on the mechanical property of the liver is relatively poor, but during the
past four years, there has been a renewed attention on soft tissue characterization due
to the development of new robotics tools and new imaging modalities. The published
materials concerning liver biomechanical properties usually include two distinct stages.
In a first stage, experimental curves relating strain and stress are obtained from specific
experimental setups and in a second stage, parameters of a known constitutive law are
fitted to these curves. Concerning the first stage, there are three different sources of
rheological data:
• ex-vivo testing where a sample of a liver is positioned inside a testing rig,
• in-vivo testing where a specific force and position sensing device is introduced
inside the abdomen to perform indentation,
TABLE 2.1
List of published articles providing some quantitative data about the biomechanical properties of the liver
Authors Experimental technique Liver origin Young modulus (kPa)
YAMASHITA and K UBOTA [1994] image-based human not available

B ROWN , ROSEN , K IM , C HANG ,
S INANAN and H ANNAFORD [2003] in-vivo porcine liver ≈ 80
C ARTER [1998] in-vivo human liver ≈ 170
D AN [1999] ex-vivo porcine liver ≈ 10
L IU and B ILSTON [2002], L IU and
B ILSTON [2000] ex-vivo bovine liver not available
N AVA , M AZZA , K LEINERMANN ,
AVIS and M C C LURE [2003] in-vivo porcine liver ≈ 90
M ILLER [2000] in-vivo porcine liver not available
S AKUMA , N ISHIMURA , K ONG
C HUI , K OBAYASHI , I NADA , C HEN
and H ISADA [2003] ex-vivo bovine liver not available
• image-based elastometry where an imaging modality like ultrasound (YAMASHITA

and K UBOTA [1994]), Magnetic Resonance Elastometry (M ANDUCA , M UTHUPIL -
LAI , ROSSMAN , G REENLEAF and E HMAN [1996]) or CT-scan (O’M AHONY,
W ILLIAMS and K ATZ [1999], H ODGSKINSON and C URREY [1992]) provides
relevant information to assess the Young modulus of living materials.
A non-comprehensive list of articles describing the liver material characteristics is
provided in Table 2.1. From this wide variety of studies, it is difficult to pick one partic-
ular constitutive model since each of experimental setup has its advantages and draw-
backs. For instance, the rich perfusion of the liver affects deeply its rheology (the liver
receives one fifth of the total blood flow at any time) and therefore it is still an open
question whether ex-vivo experiments can assess the property of living liver tissue, even
when specific care is taken to prevent the swelling or drying of the tissue. Conversely,
data obtained from in-vivo experiments should also be considered with caution because
the response may be location-dependent (linked to specific boundary conditions or non-
homogeneity of the material) and the influence of the loading tool caliper on the defor-
mation may not be well-understood. Furthermore, both the respiratory and circulatory
motions may affect in-vivo data.
Furthermore, little is known about the variability of liver characteristics between
species (does a porcine liver behave like a human liver?) but also between patients. For
instance, studies from NAVA , M AZZA , K LEINERMANN , AVIS and M C C LURE [2003]
suggest that a 20% difference in stiffness between normal and diseased livers whereas
B ROWN , ROSEN , K IM , C HANG , S INANAN and H ANNAFORD [2003] show significant
differences between in-vivo pig livers and ex-vivo cow livers.
Another important source of uncertainty in those measurements is the strain state of
the liver during indentation. Indeed, as pointed out by B ROWN , ROSEN , K IM , C HANG ,
S INANAN and H ANNAFORD [2003], most researchers precondition their liver samples
by applying several cycles of indentation in order to have more consistent estimates
of stiffness and hysteresis. However, during surgery, (rightfully) surgeons do not pre-
condition living tissues which may imply that only measurements obtained in-vivo and
in-situ through modified surgical instruments (like those developed in C ARTER [1998],
B ROWN , ROSEN , K IM , C HANG , S INANAN and H ANNAFORD [2003], NAVA , M AZZA ,
K LEINERMANN , AVIS and M C C LURE [2003]) are relevant for modeling soft tissue in
a surgical simulator.
Finally, the rheology of the liver is not only influenced by its perfusion, but also by
the Glisson’s capsule. For instance, C ARTER [1998] et al. have showed that the stiffness
of cylindrical samples of liver parenchyma with part of Glisson’s capsule is twice the
one without Glisson’s capsule, using similar rheological tests (C ARTER [1998]).
To conclude, more experimental studies are needed to reach a good understanding of
the liver biomechanical properties. Methods based on in-vivo and in-situ indentations
seem to be the most promising ones for building realistic soft tissue models in surgery
simulation. All studies demonstrate that the liver is a strongly visco-elastic material,
while L IU and B ILSTON [2002] suggest that the liver can be considered as linear elastic
for strains smaller than 0.2%.
Fortunately, in many surgical simulators, the boundary conditions governing the de-
formation of soft tissues, consist of imposed displacements only. In such case, the com-
putation of soft tissue deformation requires to solve a homogeneous system of equations
FU = 0 which is not sensitive to the absolute value of stiffness materials but to the rela-
tive stiffness between materials (G LADILIN [2002]). Hopefully, we can expect that the
relative stiffness between the liver and its neighboring organs is less variable and easier
to assess, for instance, through medical imagery.
3. Linear elastic models for surgery simulation
3.1. Main features of our approach
In the next sections, we propose three different soft tissue models that are well-suited
for the simulation of surgery and which are compatible with the constraints described
in Section 1.3. These models bear many common features that are listed below:
• volumetric structures;
• continuum mechanics based deformation;
• finite element modeling;
• linear tetrahedron finite element;
• strong approximation in dynamical modeling.
We explain the motivations of such characteristics in the next sections.
3.1.1. Using volumetric models

We can classify the geometry of anatomical structures depending on their “idealized”
dimensionality, even though they all consist of an assembly of tridimensional cells. For
instance, at a coarse scale, a blood vessel can be thought as a one-dimensional struc-
ture (Q UARTERONI , T UVERI and V ENEZIANI [2000]) whereas the gall-blader can be
represented as a two-dimensional structure (K UHNAPFEL , A KMAK and M AA [2000])
(a closed surface filled with bile). Similarly, the behavior of most parenchymatous or-
gans such as the brain, lungs, liver or kidneys are intrinsically volumetric. But one
should notice that at a fine enough scale, all anatomical structures can be considered as
volumetric.
In surgical simulators, it is frequent to rely on such dimensionality simplification in
order to speed-up computation: tubular surfaces, such as the colon, are modeled as a
deformable spline (F RANCE , L ENOIR , M ESEURE and C HAILLOU [2002]) whereas de-
formable volumetric structures, such as the liver, are represented with their surrounding
surface envelope (K UHNAPFEL , A KMAK and M AA [2000]).
However, such artifices cannot be used in a hepatic resection simulator when the
removal of hepatic parenchyma is performed.
3.1.2. Using continuum mechanics

We have chosen to rely on the theory of continuum mechanics to govern the deforma-
tion of our volumetric soft tissue models. Other alternative representations exist such as
spring–mass models (K UHNAPFEL , A KMAK and M AA [2000]), chain–mail (G IBSON ,
S AMOSKY, M OR , F YOCK , G RIMSON , K ANADE , K IKINIS , L AUER and M C K ENZIE
[1997]) or long element models (C OSTA and BALANIUK [2001]). Spring–mass models
correspond to small deformation one-dimensional elastic elements (see Section 6.1.7 for
an extended comparison) but are no longer valid for two- or three-dimensional elastic-
ity. These models are especially popular in computer graphics since they are easy to im-
plement and are based on straightforward point mechanics. The chain–mail (G IBSON ,
S AMOSKY, M OR , F YOCK , G RIMSON , K ANADE , K IKINIS , L AUER and M C K ENZIE
[1997]) is an original quasi-static deformable model based on a hexahedral mesh which
is well-suited for stiff material but does not allow any topological change. Long element
models (C OSTA and BALANIUK [2001]) correspond to valid tridimensional cylindri-
cal elastic models but are used to approximate the deformation of general volumetric
shapes.
We chose to base our soft tissue models on continuum mechanics since it offers a
well-studied and validated framework for modeling the deformation of volumetric ob-
jects unlike the methods cited above. Furthermore, it offers the following advantages:
• Scalability: when modifying the mesh topology (refinement or cutting for in-
stance), the behavior of the mesh is guaranteed to evolve continuously.
• Physical parameter identification: the elastic parameters of a biomaterial (Young
modulus, for instance) can be estimated from various methods (incremental rheo-
logical experiments, elastography or solving inverse problems). Parameter identifi-
cation for spring–mass models is known to be more difficult and requires stochastic
optimization (genetic algorithms (L OUCHET, P ROVOT and C ROCHEMORE [1995])
or simulated annealing (D EUSSEN , KOBBELT and T UCKE [1995])).
3.1.3. Using finite element modeling

Finite Element Modeling (FEM) is certainly the most popular technique for the com-
putation of structure deformation based on the elasticity theory. Furthermore, it is well-
formalized and understood and there exists many software implementations although
none of them deals with real-time deformation. Nonetheless, there exists alternative ap-
proaches such as Boundary Element Modeling (BEM) or the Finite Difference Method
(FDM).
The BEM is well-suited for the simulation of linear elastic isotropic and homoge-
neous materials (for which there exists a Green function) and is indeed a good alterna-
tive to FEM when the mesh topology is not modified. In fact, BEM has the important
advantage over FEM of not requiring the construction of a volumetric mesh. A more
thorough discussion is provided in Section 5.3.2 but this approach is not well-suited
when cutting is simulated.
The FDM is well-suited when the domain is discretized over a structured grid in
which case partial derivatives can be easily discretized. They often lead to the same
equation as FEM when specific finite elements (based on linear interpolation) are em-
ployed (BATHE [1982]). On unstructured meshes such as tetrahedral meshes, some ex-
tensions of the finite difference method have been proposed (D EBUNNE , D ESBRUN ,
C ANI and BARR [2001]) also leading to a similar equation as FEM (see discussion in
Section 4.4). With non-linear elasticity however, FEM (P ICINBONO , D ELINGETTE and
AYACHE [2003]) and FDM (D EBUNNE , D ESBRUN , C ANI and BARR [2001]) differ sig-
nificantly and no formal proof has been given that the FDM converges towards the right
solution as the mesh resolution increases.
3.1.4. Using linear tetrahedron finite element

For all finite element models described in the remainder, a simple finite element is used:
a 4-node tetrahedron with linear interpolation (P 1). The Linear Tetrahedron (LT) is
known to be a poor element (in terms of convergence) compared to the Linear Hexahe-
dron (LH) for static linear and non-linear elastic analysis (B ENZLEY, P ERRY, C LARK ,
M ERKLEY and S JAARDEMA [1995]). Also this paper shows that LH performs better
than the Quadratic Tetrahedron (10 nodes) even in a static linear elastic analysis.
The motivation for using tetrahedra rather than hexahedra clearly comes from a geo-
metrical point of view. Indeed, meshing most anatomical structures with hexahedra is
known to be a difficult task especially for structures having highly curved or circum-
voluted parts such as the liver or the brain parenchyma (Fig. 3.1). To obtain a smooth
surface envelope, it is then necessary to employ many hexahedra where a smaller num-
ber of tetrahedra would suffice. Furthermore, there exist several efficient commercial
and academic software (S IMAIL, OWEN [2000]) to fill automatically a closed triangu-
lated surface with tetrahedra of high shape quality (PARTHASARATHY, G RAICHEN and
H ATHAWAY [1993]). A second motivation for using tetrahedra rather than hexahedra
is related to the simulation of cutting soft tissue that involves removing and remeshing
of local elements. With hexahedral meshes, it is not possible to simulate general sur-
face of cut without resorting to add new element types (such as prismatic elements).
Such multi-element models (BATHE [1982]) would make the matrix assembly and local
remeshing algorithms more complex to manage.
Regarding the choice of the interpolation function (linear versus quadratic), our
choice has been mainly governed by computational issues. Given that a minimum num-
ber of tetrahedra is necessary to get a realistic visual rendering of a structure, the QT
element involves one additional node per edge compared to the LT element which on
F IG . 3.1. Example of liver meshed with hexahedra (courtesy of ESI SA).
a typical tetrahedral mesh implies at least a sixfold increase of the number of nodes.
Furthermore, we believe that the loss of accuracy in the deformation computation en-
tailed by the use of LT elements remains small compared to the large uncertainty on the
physical parameter values (Young modulus, . . . ) existing for most soft tissues.
Finally, by using linear elements, the computation of local stiffness matrices can be
done explicitly (analytically) even for non-linear elasticity. Also, the gradient of the
displacement field which is constant inside each element (constant strain) has a simple
geometric interpretation using area vectors (see Section 4.2). A significant speed-up
is therefore obtained when compared to higher order elements that require numerical
integration methods such as Gauss quadrature for estimating stiffness matrices.
3.1.5. Using large approximations of dynamic behavior

Despite the development of new in vivo rheological equipment (K AUER , V USKOVIC ,
D UAL , S ZÉKELY and BAJKA [2001]), the dynamical behavior is only known quantita-
tively for a few anatomical structures: skin, muscle, myocardium, . . . . The viscoelastic
properties of liver tissue have been studied by L IU and B ILSTON [2000] but for most
organs, constitutive laws of dynamics and their parameters must be hypothesized and
validated qualitatively.
In a surgical simulator, the boundary conditions caused by the contact with surgical
instruments can change between two iterations. Given that surgeons typically move
their instruments at low speed (typically a few millimiters per second) and making the
hypothesis that the mass of these instruments is the same or smaller than the mass
of anatomical structures, we chose to neglect the dynamics of soft tissue models in
two different ways. For a first class of models described in Section 5, we solve the
static problem F = KU (where F is the force vector, K the stiffness matrix and U the
displacement vector) at each iteration thus leading to a quasi-static approximation.
For a second class of models, described in Sections 6.1 and 7, we solve the New-
tonian equation of motion MÜ + CU̇ = −KU with the following hypotheses: the mass
matrix M is proportional to the identity matrix while the damping matrix C is diagonal.
Furthermore, in some cases, the computational time Tc is longer than the time step t
which creates a lack of synchronicity in the simulation.
3.2. Tridimensional linear elasticity
The fast computation of soft tissue deformation in a surgical simulator requires that
some hypotheses are made about the nature of the tissue material. A first hypothesis,
which leads to the two soft tissue models described in Sections 5 and 6, assumes that
soft tissue can be considered as linear elastic. The rationale behind this hypothesis is
clear: the linear relation between applied forces and node displacements leads to very
computationally efficient algorithms. But, linear elasticity is not only a convenient math-
ematical model for deformable structures: it is also a quite realistic hypothesis. Indeed,
all hyperelastic materials can be approximated by linear elastic materials when small
displacements (and therefore small deformations) are applied (F UNG [1993], M AU -
REL , W U , M AGNENAT T HALMANN and T HALMANN [1998]). It is often admitted as
reasonable to consider that a material is linear elastic when observed displacements are
less than 5% of the typical object size. In the case of hepatic tissue, a recent publication
(L IU and B ILSTON [2000]) indicates that the linear domain is only valid for strain less
than 0.2%.
Whether this constraint on the amount of displacement is valid or not in a surgical
simulator depends both on the anatomical structure and the type of surgery. For instance,
when simulating the removal of the gall bladder (cholelysectomy), the liver undertakes
small displacements but it is not the case when simulating hepatic resection or any other
surgical procedure that requires a large motion of the left lobe.
When large displacements are applied to a linear elastic material, the approximation
of hyperelasticity is no longer valid and large errors in the computation of deformation
and reaction forces can be perceived both visually and haptically. Section 7.1 describes
the shortcomings of linear elasticity in such cases.
To summarize the general equations of linear elastic materials, we proceed in four
steps. In Section 3.2.1, we provide some general definitions whereas Sections 3.2.3
and 3.2.4 give the main equations of isotropic and transversally anisotropic material.
Finally, the principle of virtual work is formulated in Section 3.2.5.
3.2.1. Definition of infinitesimal strain

We consider a three-dimensional body defined in a tridimensional Euclidian space R3 .
We describe the geometry of this body in its rest position Mrest by using material coor-
dinates X = (x, y, z)T defined over the volume of space Ω occupied by Mrest .
This body is deformed under the application of boundary conditions: these may be ei-
ther geometric boundary conditions (also called essential boundary conditions (BATHE
[1982])) or natural boundary conditions, i.e., prescribed boundary forces.
We note Mdef the body in its deformed state and Φ(x, y, z), the deformation function
that associates to each material point X located in the body at its rest position, its new
F IG . 3.2. Definition of deformation and displacement between rest and deformed positions.
position Φ(X) after the body has been deformed

φ(x, y, z),
Φ : Ω ⊂ R3 → Φ(Ω), X → Φ(X) = χ(x, y, z),
ψ(x, y, z).
The displacement vector field U is defined as the variation between the deformed posi-
tion and the rest position (see Fig. 3.2):
u(x, y, z),
U(X) : Ω → R3 , X→U= v(x, y, z),
w(x, y, z).
The observed deformation can be characterized and quantified through the analysis
of the spatial derivatives of the deformation function Φ(X). More precisely, the right
Cauchy–Green strain tensor C(X) which is a symmetric 3 × 3 matrix (therefore, has 3
real eigenvalues) is simply computed from the deformation gradient
C(X) = ∇Φ T ∇Φ. (3.1)
The Green–Lagrange strain tensor E(X), derived from the right Cauchy–Green strain
tensor, allows to analyze the deformation after rigid body motion has been removed:
1 1
E(X) = (C − I3 ) = ∇U + ∇UT + ∇UT ∇U , (3.2)
2 2
where I3 is the 3 × 3 identity matrix.
In the linear elasticity framework, applied displacements are considered as infinitesi-
mal and the Green–Lagrange strain tensor E(X) is linearized into the infinitesimal strain
tensor EL (X). This symmetric 3 × 3 tensor is simply written as

1 exx exy exz
EL (X) = [eij ] = ∇U + ∇U = exy eyy eyz .
T
(3.3)
2 e e e
xz yz zz
The diagonal elements eii of the symmetric matrix describe the relative stretch in the
direction of the reference frame, whereas off-diagonal elements eij describe shearing
quantities.
3.2.2. Definition of infinitesimal stress

The deformation of a tridimensional body is caused by applying external forces: these
forces may be either body forces FB (such as gravity forces) or surface forces FS (ap-
plied pressure) or concentrated forces FP . As a reaction to external forces, internal
forces are created inside the elastic body material.
Through Cauchy theorem (C IARLET [1987]), it is demonstrated that for each volume
element inside the deformed body, the force per unit area t(X, n) at a point X and along
the normal direction n is written as
t(X, n) = T(X)n,
where T(X) is the Cauchy stress tensor. The Cauchy stress tensor is a 3 × 3 symmetric
tensor and can be written as

σxx σxy σxz
Σ(X) = [σij ] = σxy σyy σyz .
σxz σyz σzz
The Cauchy stress Σ and infinitesimal strain EL are conjugated variables (BATHE
[1982]) which implies the following relations:
∂W ∂W
σij = , eij = , (3.4)
∂eij ∂σij
where W (X) is the amount of elastic energy per unit volume.
3.2.3. Isotropic linear elastic materials

For an isotropic linear elastic material, the elastic energy W (X) is a quadratic function
of the first two invariants of the infinitesimal strain tensor (C IARLET [1987]):
λ
W (X) = (tr EL )2 + µ tr E2L , (3.5)
2
where λ and µ are the two Lamé coefficients characterizing the material stiffness. These
two parameters are simple functions of Young modulus E and Poisson coefficients ν,
which belong to the material’s physical properties:
Eν E
λ= , µ= ,
(1 + ν)(1 − 2ν) 2(1 + ν)
µ(3λ + 2µ) λ
E= , ν= .
λ+µ 2(λ + µ)
Through Eq. (3.4), we can derive the linear relationship, known as Hooke’s law, be-
tween the stress and the infinitesimal strain tensors,
Σ = λ(tr EL )I3 + 2µEL . (3.6)
Note that the elastic energy can be written simply as a function of the linearized strain
and stress tensors,
1
W= tr(EL Σ).
2
3.2.4. Transversally anisotropic linear elastic materials

Most anatomical structures like muscles, ligaments or blood vessels are strongly
anisotropic material. This anisotropy is caused by the presence of different fibers (col-
lagen, muscle, . . . ) that are wrapped together within the same tissue. For instance,
anisotropic materials have been successfully used to model the deformation of the
eye (K AISS and L E TALLEC [1996]), of the heart (H UMPHREY and Y IN [1987],
H UMPHREY, S TRUMPF and Y IN [1990], PAPADEMETRIS , S HI , D IONE , S INUSAS ,
C ONSTABLE and D UNCAN [1999]) or the knee ligaments (W EISS , G ARDINER and
Q UAPP [1995], P USO and W EISS [1998]). In the scope of our hepatic surgery simula-
tor, we have added an anisotropic behavior where the first branches of the portal vein
are located inside the hepatic parenchyma.
We have chosen to focus only transversally anisotropic material only where there
is one direction a0 along which the material stiffness differs from the stiffness in the
orthogonal plane. Indeed, one major obstacle when modeling anisotropic material is to
get reliable data from rheological experiments regarding the directions of anisotropy
and the Young modulus in all directions. With transversal anisotropy, it is sufficient to
provide a single direction a0 and an additional pair of Lamé coefficients λa0 and µa0
(see Fig. 3.3).
The theoretical description of elastic energy of transversally anisotropic material is
largely based on the work of S PENCER [1972], S PENCER [1984] and F UNG [1993].
For the sake of clarity, we introduce the notion of direction invariant and the concept of
anisotropic stretching and shearing.
We decompose the elastic energy of a transversally anisotropic material as the sum of
the isotropic energy, provided by Eq. (3.5) and by a corrective term WAni which only
depends on the variation of Lamé coefficients:
λ = λa0 − λ, µ = µa0 − µ,
WTransv.Ani(X) = W (X) + WAni (X, λ, µ).
Without loss of generality, we can assume that a0 coincides with the z direction of
the reference frame. The isotropic elastic energy can then be written as a function of the
stretch ezz and shear (exz , eyz ) in the direction a0 :

λ 2
W (X) = + µ exx + eyy2
+ ezz2
+ λ(exx eyy + exx ezz + eyy ezz )
2
2
+ 2µ exy + eyz2
+ exz
2
.
F IG . 3.3. Definition of Lamé coefficients along the direction a0 are λa0 and µa0 .
The purpose of the corrective term WAni is to modify the isotropic Lamé coeffi-
cients in the direction of anisotropy:

λ 2
WAni (X) = − + µ ezz 2
+ λezz (tr EL ) + 2µ eyz + exz
2
.
2
The equation above can be written using the two parameters I4 and I5 which char-
acterize the strain tensor in the direction a0 (P ICINBONO , L OMBARDO , D ELINGETTE
and AYACHE [2002]):
I4 = aT0 EL a0 , (3.7)
I5 = aT0 E2L a0 . (3.8)
The first parameter I4 is simply the amount of stretch in the direction a0 whereas the to-
2 + e 2 is given by I − I 2 . With these notations, the corrective
tal amount of shearing exz yz 5 4
term can be written as

λ
WAni (X) = λI4 tr EL + 2µI5 − + µ I42 . (3.9)
2
P ICINBONO , L OMBARDO , D ELINGETTE and AYACHE [2002] proposed to decom-
pose the anisotropic term WAni (X) into a stretching and shearing part:
WAni (X) = WStr.Ani + WSh.Ani ,

λ
WSt.Ani = − + µ I42 + λI4 tr EL ,
2

WSh.Ani = 2µ I5 − I42 .
In Fig. 3.4, the distinction between stretching and shearing effects of a transversally
anisotropic material is pictured by applying a force F1 and F2 on a cylinder respectively
along and orthogonal to the direction.
3.2.5. Principle of virtual work

The equilibrium equation of a deformed body is derived through the principle of virtual
displacements. This principle states that for any compatible virtual displacement û(X)
F IG . 3.4. Comparison between isotropic (a and c) and anisotropic (b and d) cylinders (P ICINBONO ,
D ELINGETTE and AYACHE [2003]). The same horizontal (respectively vertical) loads F1 (respectively F2 )
are applied in the two leftmost (respectively rightmost) figures.
applied on a body Mdef , the total internal virtual work is equal to the total external
work. The total internal work is given by the integral of elastic energy over the body
volume whereas the external work is created by the application of body and surface
forces:

W (X) dV = ûT f B dV + ûT f S dS (3.10)
Ω Ω ∂Ω
where f B and f S
are the applied body and surface forces. Note that in Eq. (3.10), the vir-
tual displacement field û(X) is supposed to be compatible with the geometric boundary
constraints (imposed displacements). Furthermore, this relation is only valid for small
virtual displacements such that the linearized strain hypothesis still holds.
4. Finite element modeling
4.1. Linear tetrahedron element
As justified in Section 3.1.4, the computation of soft tissue deformation is based on the
finite element method. Anatomical structures of interest are spatially discretized into a
F IG . 4.1. Notations for the position and displacement vectors of a tetrahedron.
conformal tetrahedral mesh. Conformity implies that the intersection of two tetrahedra
of that mesh is either empty or consists of a vertex or an edge or a triangle.
Let Mrest be a conformal tetrahedral mesh at its rest position. The initial position of
each vertex is written as pi while its position in the deformed position is written as qi
(see Fig. 4.1). The displacement at each node is then defined as
ui = qi − pi .
We use a linear tetrahedron finite element, denoted in the literature as P1 . This amounts
to assuming a C 0 continuity of the displacement vector across the domain and equiv-
alently assuming constant strain inside each tetrahedron (since the gradient matrix is
constant inside each tetrahedron).
More precisely, if T is a tetrahedron defined by its four vertices pj , j = 0, . . . , 3, in
their rest position, then the displacement vector at a given point X = (x, y, z) ⊂ T is
defined as

3
u(X) = hj (X)uj ,
j =0
where hj (X), j = 0, . . . , 3, are the shape functions that correspond to the linear interpo-
lation inside tetrahedron T . These shape functions hj (X) correspond to the barycentric
coordinates of point X with respect to vertices pi . The analytical expression of these
shape functions is obtained from the linear relation
   px0 px1 px2 px3   
x h0
y  
y y y y
p p p p 3  h1 

 =  pz pz pz pz   h2  = PH,
0 1 2
z 0 1 2 3
1 1 1 1 1 h3
y
where pi = (pxi , pi , pzi )T are the coordinates of each tetrahedron vertex. The matrix P
completely encapsulates the shape of the tetrahedron T at its rest position. Since its
determinant |P| = 6V (T ) is the volume of T , for non-degenerate tetrahedra P can be
F IG . 4.2. Definition of area vector mi on the triangle Ti opposite to vertex pi in tetrahedron T .
inverted,
 mx m0
y
mz0 −V0 
0
−1  mx m1
y
mz1 −V1 
P−1 =  1 ,
6V (T )  mx y
m2 mz2 −V 
2 2
y
mx3 m3 mz3 −V3
where:
y
• mi = (mxi , mi , mzi )T is the ith area vector opposite to vertex pi (see description
below),
• Vi = (−1)i+1 |pi+1 , pi+2 , pi+3 | can be interpreted5 geometrically as 6 times the
volume of the tetrahedron made by the origin o and vertices pi+1 , pi+2 and pi+3 .
To simplify notations, the index i + k should be understood as (i + k) mod 4.
Area vectors mi have a very simple interpretation: they are directed along the outer
normal direction of the triangle Ti opposite to pi and their norm is equal to twice the
area of that triangle (see Fig. 4.2). More precisely, they can be computed as
mi = (−1)i+1 (pi+1 × pi+2 + pi+2 × pi+3 + pi+3 × pi+1 ), (4.1)
where pi+1 × pi+2 is the cross product between the two vectors pi+1 and pi+2 .
Because they are computed from the inverse of matrix P, these area vectors also
capture the shape of T completely, and thus play a key role when computing the stiffness
matrix K. Further properties of area vectors are described in Section 4.2.
The shape functions hi (X) can then be written as
mi · X − Vi
hi (X) = − , (4.2)
6V (T )
where mi · X is the dot product between the two vectors mi and X.
If we note that elementary volumes Vi can be expressed as
Vi = mi · pi+1 ,
5 |a, b, c| is the triple product of vectors a, b and c.
then the interpolation of displacement vectors can be written as
3
mi · (X − pi+1 )
u(X) = − ui . (4.3)
6V (T )
i=0
Finally, the interpolation matrix H(X) widely used in the finite element literature is
defined as
 
u0
 u1 
u(X) = H(X)   ,
u2
u3

h0 0 0 h1 0 0 h2 0 0 h3 0 0
H(X) = 0 h0 0 0 h1 0 0 h2 0 0 h3 0 .
0 0 h0 0 0 h1 0 0 h2 0 0 h3
4.2. Properties of area vectors
Area vectors have a major significance with respect to the geometry of a tetrahedron for
instance through the law of cosine. To write essential geometric relations, we need to
introduce the following quantities:
• normal vector ni of triangle Ti defined as the normalized area vector, ni =
mi /mi . The normal vector is pointing outward if the tetrahedron T is positively
oriented, i.e., if its volume V (T ) is positive;
• dihedral angle θi,j existing between triangle Ti and Tj and therefore between their
normal vectors ni and nj ;
• triangle area Ai , area of triangle Ti ;
• edge length li,j is the length between vertex pi and pj (see Fig. 4.3);
• foot height fi is the height of vertex pi above triangle Ti (see Fig. 4.3).
F IG . 4.3. Definition of foot height fi and edge length li,j in tetrahedron T .

The definition of area vectors gives the relation

mi = 2Ai ni .
Noting that the tetrahedron volume is simply related to the foot height and area, we get
2V (T )
mi = ni .
3fi
From the relations [P−1 ][P] = I3 and [P][P−1] = I3 , the following relations are ob-
tained:

mi = 0, (4.4)
i

mi · pi = −18V (T ), (4.5)
i

mi · pi+1 = 6V (T ), (pi+1 − pi ) · mi = 6V (T ),
i

mi · mj li,j
2
= 108V (T )2 , |mi , mi+1 , mi+2 | = (−1)i+1 36V (T )2 .
i=j, i<j
The most important result is that all area vectors sum to zero. A result of this property
is the law of cosine:
A20 = A21 + A22 + A23 − 2A1 A2 cos θ1,2 − 2A1 A3 cos θ1,3 − 2A2 A3 cos θ2,3 . (4.6)
In fact, area vectors sum to zero for any closed triangulated surface. Indeed, through
Green’s formulae (B RONSHTEIN and S EMENDYAYEV [1985]), the sum of area vectors
can be interpreted as the total flow of a constant field across a closed surface.
4.3. Computation of stiffness matrix: isotropic case
We use a displacement based finite element method which is equivalent to the classical
Ritz analysis (BATHE [1982]). On a single tetrahedron, the (linear) isotropic elastic
energy is equal to

λ
W (T ) = (tr EL )2 + µ tr E2L dV .
T 2
The gradient of the displacement ∇u(X) is constant inside T ,

3
mi · (X − pi+1 )
1 3
∇u(X) = − ∇ ui = − mi ⊗ ui ,
6V (T ) 6V (T )
i=0 i=0
where mi ⊗ ui = mi uTi is the tensor product of the two vectors,

 x x y 
mi ui mxi ui mxi uzi
 y
mi ⊗ ui = mi uxi
y y y z
mi ui mi ui .
y
mzi uxi mzi ui mzi uzi
The infinitesimal strain tensor EL is also constant inside T :

−1
3
1
EL = ∇u + ∇uT = (mi ⊗ ui + ui ⊗ mi ).
2 12V (T )
i=0
The first invariant (tr EL )2 is simply

3 2
1
1
3
(tr EL )2 = 2
mi · ui = uTi [mi ⊗ mj ]uj .
144V (T ) 144V (T )2
i i,j =0
The second invariant is slightly more complex to obtain:
3
1
tr EL =
2
2
tr (mi ⊗ uj )(ui · mj ) + (ui ⊗ mj )(mi · uj )
144V (T )
i,j =0

+ (mi ⊗ mj )(ui · uj ) + (ui ⊗ uj )(mi · mj )
3

= 2
uTi (mj ⊗ mi ) + (mj · mi )I3 uj .
72V (T )
i,j =0
Finally, the linear elastic energy is a quadratic function of the displacement and is writ-
ten as
1
3

W (T ) = uTi λ(mi ⊗ mj ) + µ(mj ⊗ mi ) + µ(mi · mj )I3 uj ,
72V (T )
i,j =0
3

W (T ) = uTi BTij uj , (4.7)
2
i,j =0
where [BTij ] is the 3 × 3 stiffness matrix of tetrahedron T between vertices i and j .

Noting that [mi ⊗ mj ]a = mi (mj · a), we can write the local elastic energy in terms of
dot products,
1
3

W (T ) = λ(ui · mi )(mj · uj ) + µ(ui · mj )(mi · uj )
72V (T )
i,j =0

+ µ(mi · mj )(ui · uj ) .
Since mi ⊗ mj = (mj ⊗ mi )T , it is clear that local tensors are symmetric matrices:
[BTij ] = [BTj i ]T . Therefore, there are only 10 distinct local stiffness matrices with four
vertex matrices [BTii ] and six edge matrices [BTij ], i = j .
4.3.1. Local vertex stiffness matrix

Vertex stiffness matrices take the simple form with normal vector ni :
T A2i
Bii = (λ + µ)(ni ⊗ ni ) + µI3 . (4.8)
9V (T )
These matrices have eigenvalues (λ + 2µ, µ, µ), ni being the first eigenvector, and
two directions orthogonal to ni being the two other eigenvectors.
4.3.2. Local edge stiffness matrix

The stiffness matrix between vertex i and j is
T 1
Bij = λ(mi ⊗ mj ) + µ(mj ⊗ mi ) + µ(mi · mj )I3 . (4.9)
36V (T )
This matrix has the edge direction (pj − pi )/pj − pi as first eigenvector associated
with eigenvalue (µ(mi · mj ))/(36V (T )). The existence of the other two eigenvectors
depends on the sign of the following matrix determinant:

(λ + µ)(mi · mj ) µmi 2 λ
= λµAi Av 2 1 +
2 2
cos θi,j − 1 .
2
λmj 2 2µ(mi · mj ) µ
4.3.3. Global stiffness matrix

The elastic energy of the whole deformed body is then computed by summing Eq. (4.7)
over all tetrahedra. This total energy W (Mdef ) may be written with the displacement
vector U, gathering all displacement vectors ui , and a global stiffness matrix K:
1
W (Mdef ) = UT KU. (4.10)
2
This stiffness matrix K is built by assembling local stiffness matrices [BTij ]. Because
these local matrices are symmetric with respect to the swap of indices [BTij ] = [BTj i ],
the global stiffness matrix K is symmetric.
4.3.4. Global vertex stiffness matrix

The 3 × 3 submatrix [Ki,j ] associated with vertices i (row index) and j (column index)
is computed as the sum of local stiffness matrices for all tetrahedra containing both
vertices. The set of tetrahedra adjacent to a given vertex (respectively edge) is called the
shell S(i) of this vertex (respectively edge). In particular, for diagonal submatrices, we
get

1
[Ki,i ] = (λT + µT )(mi ⊗ mi ) + µT A2i I3 .
36V (T )
T ∈S (i)
In fact, we can provide a rather simple interpretation of this matrix expression. Its
first term can be seen as the inertial matrix (second order moment) of area vectors mi
weighted by (λT + µT )/(72V (T )) (see Fig. 4.4). Indeed, on a manifold tetrahedral
mesh (but not on all conformal tetrahedral meshes), the shell of an interior vertex is
homeomorphic to a sphere and it can be easily proved that the sum of its area vectors
is null (Minkowsky’s sum). Therefore mi ⊗ mi represents the local contribution to an
inertia matrix. If vertex pi is surrounded by semi-regular tetrahedra, then the matrix of
inertia is proportional to identity. Note also that because it is weighted by the inverse
of the tetrahedron’s volume, it is very sensitive to the disparity in tetrahedra size. The
second part is simply the sum of the second Lamé coefficient weighted with the inverse
of the tetrahedron volume.
F IG . 4.4. Shell of a vertex pi in a tetrahedral mesh: only the opposite triangle is drawn. For each triangle the
area vector mi is pointing outward. The sum of these area vectors is null and their weighted matrix of inertia
determines the global stiffness matrix at pi .
4.3.5. Global edge stiffness matrix

The off-diagonal terms [Ki,j ] of the stiffness matrix correspond to edge stiffness
matrices that are the sum of local edge stiffness matrices. The edge direction
(pj − pi )/pj − pi is an eigenvector of [Ki,j ] associated with the eigenvalue ki,j :

µT (mi · mj )
ki,j = ,
36V (T )
T ∈S (i,j )
where S(i, j ) is the set of tetrahedra adjacent to that edge (its shell). The tetrahedron
volume V (T ) can be written as a function of triangle areas,
2
V (T ) = opp Ai Aj sin θi,j ,
3li,j
F IG . 4.5. Shell of an edge linking vertices pi and pj in a tetrahedral mesh: the adjacent tetrahedra are drawn
with dashed lines whereas opposite edges to that edge are drawn with solid lines. One of the eigenvalues of
opp
edge stiffness matrix depends on the weighted sum of the lengths li,j .
opp
where li,j is the length of the opposite edge in tetrahedron T (see Fig. 4.5).
1
opp
ki,j = µT li,j cot θi,j .
6
T ∈S (i,j )
4.4. Physical interpretation of isotropic stiffness matrix
Eq. (4.7) describes the local stiffness matrices and it can be interpreted as the sum of
discrete differential operators. Indeed, the isotropic elastic energy related to the first
invariant of the infinitesimal strain tensors, can be written as a quadratic functional of
the displacement vector:
λ λ µ
W (X) = (tr EL )2 + µ tr E2L = (div u)2 + µ tr ∇uT ∇u − curl u2 .
2 2 2
The first variation of the elastic force −δW can be interpreted as the density of linear
elastic force per unit volume, and is given by the Lamé equation:
−δW = (λ + µ)∇(div u) + µu.

It is therefore natural to compare the Lamé equation with the expression of the dis-
crete elastic force Fi (T ) acting on a vertex i of tetrahedron T :
3
T −1

3
Fi (T ) = − Bij uj = λ(mi ⊗ mj ) + µ(mj ⊗ mi )
36V (T )
j =0 j =0

+ µ(mi · mj )I3 uj .
The three terms of the local rigidity matrix may be interpreted as follows:
−1

3

Fi (T ) = λ(mi ⊗ mj ) + µ(mj ⊗ mi ) + µ(mi · mj )I3 uj .
36V (T )
j =0
T1 T2 T3
∇(div u) ∇(div u) u
operator pseudo-operator operator
The first term of the local rigidity matrix corresponds to the integral of the operator
∇(div u) over a subdomain of tetrahedron T . Indeed, through Green’s second formu-
lae (B RONSHTEIN and S EMENDYAYEV [1985]), the integral over a domain D of that
operator can be evaluated along its boundary ∂D:

∇(div u) = (div u)n dS.
D ∂D
However, the divergence operator is actually constant over T and is equal to
−1
3
mj · uj .
6V (T )
j =0

Furthermore, since the integral n dS over each triangle of T is equal to 12 mi , T1 is
equal to one third of the flux of (div un) through the face of T opposite to vertex pi :
3
−1

3
1 1 mi
T1 = (mi ⊗ mj ) uj = mj · uj .
36V (T ) 3 6V (T ) 2
j =0 j =0
Thus, we can provide a straightforward interpretation on the first term of the local rigid-
ity tensor: it corresponds to the integration of the ∇(div u) operator over a subdomain
Di for which

mi
n dS = .
T Di 6
A natural choice for this subdomain√is to consider the shell Si of vertex pi homotheti-
cally scaled down by a ratio of 1/ 3. This subdomain is sketched √ in Fig. 4.6(a) and
(b): the vertices of the subdomain are located at distance
√ of 1/ 3 times the original
edge length from vertex pi . Unfortunately, since 1/ 3 > 0.5, the subdomain of two
neighboring vertices overlap.
To obtain a non-overlapping subdomain Di , one should consider the subdomain de-
fined by the middle of each edge, the barycenters of each triangles and the barycen-
ter of the tetrahedron, as proposed by P UTTI and C ORDES [1998]. More precisely, as

F IG . 4.6. Definitions of two subdomains for which n dS is equal to one third the value through triangle i ,
opposite of vertex pi in tetrahedron T ; (a) and (b): front and side view√of the first subdomain consisting of
a single triangle corresponding to the homothety i with a ratio of 1/ 3; (c) and (d): front and side view
of the second non-overlapping subdomain consisting of 6 triangles linking the edge middles, triangle centers
and tetrahedron center.
shown in Fig. 4.6(c), the subdomain consists in the six triangles (FAG, GAB, BGC,
CGD, DGE, EGF ) where A, C, E are the centers of the three triangles adjacent to pi ,
B, D, F are the centers of the three adjacent edges and G is the tetrahedron barycenter.
This subdomain is called the barycentric dual cell in C OSMI [2001].
Indeed, the flux over the six triangles may be written as a sum of cross products,

n dS = A × B + B × C + C × D + D × E + E × F.
Di
Since A, B, C, D, E, F, G are simple barycentric coordinates of the four tetrahedron
vertices pi , pj , pk , pl , it can be simply evaluated as a function of these vertices,

1 mi
n dS = (pj × pk + pk × pl + pl × pj ) = .
Di 6 6
Thus, to summarize, we have proved so far that term T1 is the integral of the ∇(div u)
operator over a non-overlapping subdomain centered on pi .
The second term T2 of the local rigidity matrix is the transposed matrix of the first
term T1 but cannot be interpreted in terms of a linear differential operator. In fact, if we
write T2 as (mi · uj )mj we can state that T2 corresponds to the flux of a scalar field
equal to 12V1(T ) (mi · uj ) over each face of the subdomain Di . It should be noticed that
T2 has no equivalent in the continuous formulation (the Lamé equation) and is produced
by the evaluation of curl u2 .
The third term T3 corresponds to the discrete Laplacian operator and its expression
originates from the evaluation of 12 tr(∇uT ∇u). The same approach as for the ∇(div u)
can be applied. First, the integral of the Laplacian operator is integrated over a domain
D using the integral Gauss theorem. For the x component ux of the displacement field,
it gives

x x
u dV =
x
∇ · ∇u dV = ∇u · n dS.
D D ∂D
If the domain D is included inside a tetrahedron, then the gradient of the displacement
field is a constant vector,
−1
3
∇ux = mj uxj .
6V (T )
j =0
If we suppose that the domain boundary coincides with triangle i , opposite to pi in

tetrahedron T , then we get

−1
1
3 3
mi
ux dV = mj mj uxj · .
D 6V (T ) 6V (T ) 2
j =0 j =0
Therefore, T3 corresponds to the integral of the Laplacian operator over a domain D for
which

mi
n dS = ,
∂ DT 2
for instance the subdomain defined in Fig. 4.6(c), corresponding to the barycentric dual
cell of vertex pi in tetrahedron T . The Finite Element approximation of the Lapla-
cian operator on tetrahedra was previously studied by P UTTI and C ORDES [1998], D E -
BUNNE , D ESBRUN , C ANI and BARR [2001] and C OSMI [2001].
To summarize, we have proved that the variational formulation of linear elasticity
over tetrahedral meshes is not completely equivalent to the Finite Difference and Finite
Volume methods. Indeed, the latter methods are equivalent to the differential formu-
lation of Finite Element method which leads to the following equation of the elastic
force:
−1

3

Fi (T ) = (λ + µ)(mi ⊗ mj ) + µ(mi · mj )I3 uj . (4.11)
36V (T )
j =0
The variational formulation of the FEM creates the stiffness matrix from the expression
of the elastic energy whereas the differential formulation of the FEM is based on the
Lamé differential equation.
4.5. Computation of stiffness matrix: transversally anisotropic case
From Section 3.2.4, the density of elastic energy for a transversally anisotropic material
can be derived from the isotropic case by adding a corrective term:
W (X)Transv.Ani = W (X) + WAni (X),

λ
W (X)Transv.Ani = W (X) + λI1 I4 + 2µ I5 − + µ I42 ,
2
where λ and µ are the variation of Lamé coefficient in the direction of anisotropy
a0 and where I4 and I5 are the constants defined in Eqs. (3.7) and (3.8). The evaluation
of I4 and I5 with the linear tetrahedron finite element gives:
−1
3
I4 = (a0 · mi )(a0 · ui ),
6V (T )
i=0
1
(tr EL )I4 = 2
uTi (a0 · mj )(mi ⊗ a0 ) uj ,
72V (T )
1
3

I42 = 2
uTi (a0 · mi )(a0 · mj )(a0 ⊗ a0 ) uj .
36V (T )
i,j =0
Similarly for I5 :
3

I5 = 2
uTi (a0 · mj )(a0 ⊗ mi ) + (a0 · mi )(mj ⊗ a0 )
144V (T )
i,j =0

+ (mi · mj )(a0 ⊗ a0 ) + (a0 · mi )(a0 · mj )I3 uj .
Thus the additional elastic energy term W (T )Ani due to transversal anisotropy can
also be written as a bilinear function of vertex displacements,
1
T T
3
W (T )Ani = ui Aij uj (4.12)
2
i,j =0
with the local 3 × 3 matrix [ATij ] being defined as

T 1
Aij = λ(a0 · mj )(mi ⊗ a0 )
144V (T )
− (λ + 2µ)(a0 · mi )(a0 · mj )(a0 ⊗ a0 )
+ µ(a0 · mj )(a0 ⊗ mi ) + µ(a0 · mi )(mj ⊗ a0 )

+ µ(mi · mj )(a0 ⊗ a0 ) + µ(a0 · mi )(a0 · mj )I3 .
4.5.1. Local vertex stiffness matrix

When i = j , the vertex stiffness matrix is written as
T 1
Aii = (λ + µ)(a0 · mi )(mi ⊗ a0 )
144V (T )
− (λ + 2µ)(a0 · mi )2 (a0 ⊗ a0 )
+ µ(a0 · mi )(a0 ⊗ mi ) + µmi 2 (a0 ⊗ a0 )

+ µ(a0 · mi )2 I3 .
This matrix has c0 , the unit vector orthogonal to both a0 and mi as first eigenvector
with eigenvalue µ(a0 · mi )2 . The existence of the other two eigenvectors, in the plane
defined by a0 and mi , depends on the sign of (2µ + λ)(a0 · mi )2 − λmi 2 .
4.5.2. Global stiffness matrix

For a transversally anisotropic material, the global stiffness matrix K is assembled as
the sum of local isotropic and anisotropic stiffness matrices:

[Ki,j ] = [Bi,j ] + [Ai,j ]. (4.13)

T ∈S (i,j )
One should note that the global matrix [Ki,j ] contains non-null values only if vertices i
and j are linked by an edge of the tetrahedral mesh.
4.6. Work of gravity forces
We calculate the potential energy of gravity forces when a displacement field u(X) is
applied on the body Mdef . If we write g the gravity vector (g = 9.8 m/s2 ) and ρ the
density of the material (assumed constant for the whole body), then the potential energy
of a tetrahedron T is a simple function of the center of mass T :

q0 + q1 + q2 + q3
Wg (T ) = ρX · g dV = ρ X dV · g = ρV (T ) · g.
T T 4
If we drop the constant part of this energy, which is equivalent to consider the work of
gravity forces when a displacement field u(X) is applied, then we get
u0 + u1 + u2 + u3
Wg (T ) = ρV (T ) ·g
4
 
1
ρV (T )g 1
= [ u0 u1 u2 u3 ]   .
4 1
1
The potential energy of the whole model Mdef is the dot product of the following two
vectors:

...
g
Wg = Wg (T ) = U R = U ri ,
T g T
T ...
g
where Rg is a vector of size 3N . More precisely, the sub-vector ri of Rg associated
with vertex i is proportional to the gravity vector, the coefficient being the volume of its
neighboring tetrahedra:

g V (T )
ri = ρ g. (4.14)
4
T ∈S (i)
4.7. Work of external surface pressure
Among external forces acting on deformable soft tissue models, we include a pressure
force fp which is applied on a part of its surface. We consider that such pressure force
has a constant intensity fp = p but its direction may be either constant (contact with
a stream of gas) or directed along the surface normal (contact with a solid, fluid or gas
at low speed). In the latter case, the force applied on a triangle T is
fp (T ) = p n(T ).
For a tetrahedral mesh, we consider that such constant pressure fp is applied on a set C
of surface triangles. If we consider a triangle T ∈ C consisting of vertices (pi , pj , pk ),
the work of fp on this triangle is

ui + uj + uk
Wp (T ) = fp · u(X) dA = A(T ) fp · .
T 3
The work of external surface pressure on the whole model Mdef is then

...
p
Wg = UT Rp = UT ri (4.15)
...
p
where ri is null if vertex pi is not adjacent to any triangles in C and is proportional the
sum of triangles area otherwise:
p

A(T )fp (T )
ri = .
3
pi ∈T , T ∈C
F IG . 4.7. The pressure applied on neighboring triangles results in a force directed along the surface normal
at a vertex and proportional to the sum of neighboring triangle area. The vertex surface normal ni is computed
as the weighted average of triangle normals.
p
If the pressure force is applied along the surface normal, then vector ri has an intuitive
formulation. The nodal force, resulting from the pressure applied on neighboring trian-
gle, is proportional to the area sum of surrounding triangles and is directed along the
surface normal ni at vertex pi (see Fig. 4.7):

p p
ri = A(T ) ni ,
3
pi ∈T , T ∈C
where ni is computed as the average of surrounding triangle normals n(T ) weighted by

their area,

T ∈C n(T )A(T )
ni = .
T ∈C A(T )
4.8. Mass matrix
The mass matrix is derived from the evaluation of the kinetic energy E(Mdef ) on the
whole body Mdef . The density of kinetic energy w(X) = ρ(u̇(X))2 where u̇(X) =
du/dt is the speed of the material point X. It follows that the kinetic energy of tetrahe-
dron T is a bilinear form of the speed of nodal vertices u̇i :
 T T T T 
 T M0,0 M0,1 M0,2 M0,3  
U̇0 U̇0
 T T T T 
 U̇1   M1,0 M1,1 M1,2 M1,3   U̇1 
E(T ) =    T  .
U̇2  M2,0 MT2,1 MT2,2 MT2,3  U̇2
U̇3 U̇3
MT3,0 MT3,1 MT3,2 MT3,3
This tetrahedron mass matrix has size 12 × 12 and is composed of 4 × 4 local mass
matrix between vertex i and j , MTi,j that are 3 × 3 diagonal matrices,

MTi,j = ρ hi (X)hj (X) dV I3 .
T
To evaluate the integral, we use the 3 barycentric coordinates (h0 , h1 , h2 ) as integration

variables. Based on Eqs. (4.3) and (4.4), the determinant of the Jacobian matrix is equal
to the inverse of 6V (T ),

∂h0 ∂h1 ∂h2 1 1

∂X ∂X ∂X = 216V (T )3 |m0 m1 m2 | = 6V (T ) .
Thus the integral can be computed explicitly using the expression below:
1 1−h0 1−h0 −h1
hi (X)hj (X) dV = 6V (T ) hi hj dh0 dh1 dh2
T 0 0 0
V
= if i = j
10
V
= if i = j.
20
Thus the local mass matrix MTi,j is equal to ρV10(T ) I3 if i = j and to ρV20(T ) I3 , otherwise.
If we perform mass lumping by considering only diagonal elements equal to the sum of
row values, then we naturally get ρV 4(T ) I3 , as if the tetrahedron mass is evenly spread
over its four vertices.
The kinetic energy of the whole body can be written as a function of the global mass
matrix built by assembling the local matrices MTi,j ,
1 1
E(Mdef ) = U̇T MU̇ = U̇T [Mi,j ]U̇,
2 2
where Mi,j , the global 3 × 3 mass matrix between vertex i and j , depends on the
volumes of tetrahedra adjacent to vertex i (if i = j ) or tetrahedra adjacent to edge (i, j )
if i = j :

V (T )
Mi,i = ρ I3 , (4.16)
10
T ∈S (i)

V (T )
Mi,j = ρ I3 if i = j. (4.17)
20
T ∈S (i,j )
If we perform mass lumping to get a diagonal mass matrix M (and therefore eas-
ily invertible), then the vertex mass is equal to one fourth of the mass of its adjacent
tetrahedra:

V (T )
(Mi,i )lumping = ρ I3 .
4
T ∈S (i)
4.9. Boundary conditions
In a surgical simulator, the boundary conditions of a soft tissue model are related to the
existence of contacts with either its neighboring anatomical structures or with surgical
tools (Fig. 4.8).
F IG . 4.8. The three different boundary conditions resulting from interaction with neighboring structures or
with surgical tools.
We simplify the interaction with other physical material by considering that such an
interaction can be represented either in terms of imposed displacements or elastic forces
or surface pressure forces. If the material is stiff, or if it is significantly stiffer than the
material of interest, we model the contact by imposing given displacements on a set of
vertices. For instance, in the case of the liver model, we consider that vertices located
near the vena cava (a stiff vessel) are stable (zero displacement).
If neighboring materials are as stiff (or less) than the material of interest, then we
model the interaction as a linearized spring force. More precisely, for a boundary ver-
tex pi , the applied force rei is directed along a given direction d, with stiffness ke and
rest displacement uei :

rei = −ke ui − uei · d d = −ke (d ⊗ d) ui − uei . (4.18)
Using a linearized spring allows to compute the static equilibrium by solving a linear
system of equation. Indeed, the stiffness caused by the spring ke (d ⊗ d) can be added
to the global stiffness matrix while the residual force ke (d ⊗ d)uei is added to the nodal
load at node i. Furthermore, since the stiffness ke is lower than the Young modulus of
the material, the condition number of the updated stiffness matrix is not significantly
modified.
In the sequel, we do not consider linearized spring boundary conditions explicitly.
Instead, we modify the global stiffness matrix K into K , and we consider that a nodal
force rbi is applied to vertex pi ,

Ki,i = Ki,i + ke (d ⊗ d) , rbi = ke (d ⊗ d)uei .
When a soft tissue model is in contact with some fluids (bile, water, blood, . . .) or
gas (carbon dioxide, air, . . .), we make the hypothesis that a constant pressure is applied
along the normal direction of the contact surface. The computation of the nodal forces
is detailed in Section 4.7.
Finally, the contact between surgical tools and a soft tissue model may be posed, in
theory, either as imposed displacements (geometric method (BATHE [1982])) or as pre-
scribed forces (penalty method (BATHE [1982])). However, in practice, the motion of
surgical tools is controlled by the end-user through a force-feedback device. To decrease
their cost, these devices are force-controlled and follow a simple open loop: the posi-
tions of surgical tools can be sent to a computer while they receive the force level that
should be felt by the end-user. In other words, despite the low speed of a surgeon hands
the position of a surgical tool varies significantly between two iterations (dt = 20 ms)
and therefore we found that the penalty method was not suited for deforming a soft
tissue model.
Thus, after detecting the collision between soft tissue models and surgical tools, a set
of imposed displacements at the collision nodes is computed. This computation is ob-
viously ill-posed since it relies only on geometry (surface–volume intersection) rather
than physical principles (Coulomb friction, for instance). Furthermore, a major chal-
lenge is to design a stable contact algorithm where a small displacement of the tool
entails a small variation of node position. The geometric contact algorithm used in our
hepatic surgery simulator, can be found in P ICINBONO , L OMBARDO , D ELINGETTE
and AYACHE [2002].
To summarize, we consider only 2 types of boundary conditions in the remainder:
(1) Imposed displacement. We write Vd the set of vertices pi for which the displace-
ment ubi is known. In the scope of surgery simulation, these vertices are always
lying on the surface of the mesh.
(2) Applied nodal forces. We write Vf the set of vertices pi where an external force
rib is applied. Again, we make the hypothesis that applied forces may exist only
on surface nodes.
4.10. Equilibrium equations
We apply the principle of virtual displacements described in Section 3.2.5 to obtain the
finite element formulation of equilibrium equations. In a first stage, we only consider
the static equilibrium by neglecting inertial forces. Thus, based on Eq. (3.10), we can
state that the virtual elastic energy is equal to the sum of the virtual work of gravity and
boundary forces,
1 T T g T b
U KU = U R + U R .
2
Since this equation must hold for any set of compatible displacements, the static equa-
tion of equilibrium becomes
KU = Rg + Rb . (4.19)
It is important to note that Eq. (4.19) is written for all nodes including the Vd
nodes where the displacement is imposed. Therefore, in order to compute the un-
known displacement vectors (where no displacement is imposed), it is important to
write Eq. (4.19) with a distinction between free nodes (subscript f ) and constrained
nodes (subscript c):
! ! g !
Kff Kf c Uf Rf + Rbf
= g
Kcf Kcc Uc Rc + Rbc
thus leading to
g
Kff Uf = Rf + Rbf − Kf c Uc . (4.20)
In the case of a linear tetrahedron finite element, Kf c Uc is non-zero only for free nodes
that are neighbors to fixed nodes. In the remainder, we used simplified notations by
dropping the subscript f for the stiffness matrix and displacement vector and by gath-
ering all applied nodes into a single vector:
KU = R. (4.21)
To get the dynamic law of motion, the work of inertial forces − 12 U̇T MU̇
should be
added to the work of body forces. By adding the work of damping forces, the following
classical equation is obtained:
MÜ + CU̇ + KU = R, (4.22)

where C is the damping matrix. In general, we assume that C follows Rayleigh damp-
ing,
C = γ1 M + γ2 K. (4.23)
This assumption is important for performing modal analysis but also for ensuring that
the damping matrix, as the stiffness matrix, is also sparse.
4.11. Solution of equilibrium equations
The static equilibrium given by Eq. (4.21) is a linear system of equations with a sym-
metric positive definite stiffness matrix. Since this matrix is sparse, the classical method
to solve this equation is to use the conjugated gradient algorithm [S AAD , 1996].
More precisely, when solving the complete system KU = R, we perform the follow-
ing steps:
• Node renumbering by using the reverse cutting McKee algorithm (S AAD [1996])
in order to decrease the bandwidth of the stiffness matrix.
• Matrix preconditioning based on Cholesky factorisation or incomplete LU decom-
position (S AAD [1996]).
• Application of the conjugated gradient algorithm for solving the linear system of
equation. We rely on the Matrix Template Library (L UMSDAINE and S IEK [1998])
for an efficient implementation of these algorithms in C++. When the stiffness
matrix is poorly conditioned, for instance, for nearly incompressible materials, it
is possible that the conjugated gradient algorithm fails. In which case, we resort
to using direct methods for solving the system of equation, such as Gauss pivoting
(S AAD [1996]).
Despite optimizing the bandwidth and the condition number of the stiffness matrix, the
time required for solving the static equation is still too large for real-time computation.
For instance, with a liver model composed of a mesh consisting of 1313 vertices, the
solution of the linear system of size 3939 × 3939 requires 9 s on a PC Pentium II
(450 MHz) with 140 iterations of the preconditioned conjugated gradient in order to
reach an accuracy of 0.001 mm.
Therefore, solving directly the static equation with the conjugated gradient al-
gorithm does not satisfy the real-time constraints mentioned in Section 1.3.2 since
Tc > Trelaxation . As an alternative, we propose in the next sections, three soft tissue mod-
els that satisfy either hard or soft real-time constraints.
5. Quasi-static precomputed linear elastic model
5.1. Introduction
Since the complete solution of the static equilibrium equation is too computationally
expensive for real-time constraint, a straightforward solution is to perform only few it-
erations of the conjugated gradient at each time step in order to increase the update rate.
This approach, proposed by BARAFF and W ITKIN [1998] is well-suited in the context
of computer animation but is not applicable for surgery simulation where boundary con-
ditions are constantly changing and are formulated in terms of imposed displacements.
Indeed, using a conjugated gradient method would require to modify the stiffness ma-
trix frequently as well as its preconditioning which would considerably reduce its effi-
ciency.
Instead, we propose a quasi-static precomputed linear elastic model (C OTIN ,
D ELINGETTE and AYACHE [1999a]) that is based on a simple concept which consists in
partially inverting the stiffness matrix in a precomputation stage before the simulation.
This model has the following characteristics:
• It is computationally very efficient: the computation complexity during the simu-
lation is proportional to the cube of the number of imposed displacements.
• Only the position of surface nodes is updated during the simulation. In fact, only
the data structure of the triangulated surface corresponding to the shell of the tetra-
hedral mesh is needed online.
• During the simulation the reaction forces at the nodes where the virtual instruments
collide are also computed.
• The model is quasi-static, i.e., it computes the static equilibrium position at each
iteration.
However, it relies on the following hypotheses:
• The mesh topology is not modified during the simulation. Thus, no simulation of
cutting or suturing can be performed on this model.
• The interaction with neighboring tissues or with instruments is translated into mod-
ified boundary conditions (displacements or forces) only on surface nodes but not
on the boundary conditions of internal nodes.
Therefore, the main limitation of this precomputed model comes from the first hypoth-
esis which states that it is not suited for the simulation of tissue cutting.
5.2. Overview of the algorithm
One important feature of the model consists in making a distinction between surface
and interior nodes. Thus, for the sake of clarity, we decompose the displacement and
load vectors as well as the stiffness matrix according to surface and interior nodes with
the s and i subscripts:
! ! !
Kss Ksi Us Rs
= .
Kis Kii Ui Ri
It is important to note that only free vertices appear in this matrix as discussed in Sec-
tion 4.10.
The solution of static equation can be obtained by multiplying the compliance ma-
trix [G], corresponding to the inverse of the stiffness matrix [K], with the load vector.
This compliance matrix can also be decomposed into surface and interior nodes,
! ! !
Us Gss Gsi Rs
= . (5.1)
Ui Gis Gii Ri
The load vector Rs that applies on free surface nodes can be decomposed into two
parts. A first part R0s , corresponds to loads that will not evolve during the simulation for
instance gravity forces (see Section 4.6), constant pressure forces (see Section 4.7), ap-
plied nodal forces (see Section 4.9) or the presence of a non-zero imposed displacement
vertex in its neighborhood (see Eq. (4.20)). The second part RC s corresponds to loads
that are created by the contact of the soft tissue with surgical tools.
The principle of this soft tissue model is to compute the surface node positions Us
directly from the contact loads RC s by multiplying this vector with the compliance ma-
trix Gss :
Us = Gss RC
s + Us ,
0
(5.2)
U0s = Gss R0s + Gsi Ri .
Since the loads on interior nodes Ri do not evolve during the simulation, U0s is a dis-
placement offset that is computed as the displacement of surface nodes when no contact
s = 0.
loads are applied: RC
The goal of the precomputation stage is to compute the compliance matrix [Gss ].
5.3. Precomputation stage
5.3.1. Description of the algorithm

ij
In the remainder, we write [Gss ] the 3 × 3 submatrix of Gss associated to vertex i
j
and j . More precisely, a force Rs applied on vertex j entails an additional displacement
ij j
of vertex i equal to [Gss ]Rs .
The algorithm for computing the compliance matrix Gss is described as Algorithm 1.
It consists in solving 3 × Ns times the linear system of equations KU = R, where Ns
is the number of surface vertices. Note that the size of the stiffness matrix K is N =
Ns + Ni whereas the size of the compliance matrix Gss is 3Ns × 3Ns .
1: Set Ri = 0
2: for all Surface Vertex i do
3: for all j such that 0 j 2 do
4: Set Rs = 0
5: Set to 1.0 the j th component of the load Ris applied to vertex i
6: Solve the static equilibrium equation KU = R
7: for all Surface Vertex k do
8: Store the computed displacement Uk of vertex k into the j th column
of matrix [Gkiss ]
9: end for
10: end for
11: end for
A LGORITHM 1. Computation of the compliance matrix Gss .
The solution of equation KU = R is performed using the steps described in Sec-

tion 4.11 including node renumbering and matrix preconditioning. Since the rigidity
matrix K is the same for all 3 × Ns systems of equations, these two steps are performed
only once, which significantly speeds-up the computation. Each time a linear system
of equation is solved, the displacement of all surface nodes Us corresponds to a col-
umn of matrix Gss . The storage of matrix Gss requires only (8 × 9(Ns )2 )/2 bytes (each
element being stored as a double), since it is a symmetric matrix, as the inverse of a
symmetric matrix.
Algorithm 1 can be slightly improved in the following way:
• Applying a unitary force successively along the X, Y and Z directions may cause
a loss of accuracy in computing the compliance matrix, because the resulting dis-
placement may be very large or very small depending on the size of the mesh. To
obtain meaningful displacements, it is possible to apply a force fref and then di-
vide the resulting displacement by fref to compute Gss . A good choice for fref is
[Ki,i ] ∗ 0.1 ∗ l, where [Ki,i ] is the block diagonal stiffness matrix of vertex i,
and l is the estimated diameter of the object. This choice of force scale, produces
displacements which are roughly equal to 10% of the diameter.
• It is sometimes necessary to obtain the displacement of some interior nodes during
the simulation. This is the case, for instance, when vessels or tumors, located inside
an organ, need to be displayed during the simulation. In this case, it is possible in
the final loop of the algorithm (lines 7, 8 and 9 of Algorithm 1) to add these inside
vertices to the list of surface vertices. Thus, it does not entail the solution of any
additional system of equations, but only an additional storage requirement since
the compliance matrix becomes a rectangular matrix of size 3Ns × 3(Ns + Ni )
where Ni is the number of additional interior nodes.
This precomputation stage is quite computationally expensive and requires between a
few minutes up to several hours depending on the number of the mesh vertices and the
stiffness of the material. For instance, the liver model presented in Fig. 5.1 is composed
of 1394 vertices, 8347 edges and 6342 tetrahedra. Its triangulated surface is composed
of 1224 triangles and 614 vertices which is enough to produce a smooth visual ren-
F IG . 5.1. Visualization of a liver model with 1394 vertices and 6342 tetrahedra.
dering. The Poisson ratio of the material is set to 0.45 while its Young modulus is
E = 1000 kPa. In this case, the precomputation time required nearly 4 h on a Pentium
PII 450 MHz, while the compliance matrix is stored in a file of size 13 Mb.
5.3.2. Other methods for computing the compliance matrix

At least two alternative methods have been proposed in the literature to compute the
compliance matrix Gss . The first one, proposed by B RO -N IELSEN and C OTIN [1996] is
based on matrix condensation (M ACMILLAN [1955]). More precisely, the compliance
matrix Gss can be directly obtained from the inversion of the stiffness matrix Kii of
interior nodes. From Eq. (5.1), we can derive the following equations:
Kii Ui = Ri − Kis Us ,

Kss Us + Ksi K−1 −1
ii Ri − Kii Kis Us = Rs ,

Kss − Ksi K−1 −1
ii Kis Us = Rs − Ksi Kii Ri . (5.3)
From Eq. (5.3), we can deduce the expression of the compliance matrix,
−1 −1
Gss = K ss = Kss − Ksi K−1ii Kis . (5.4)
Therefore, the computation of Gss requires the inversion of two matrices: the first one of
size 3Ni × 3Ni and the second one of size 3Ns × 3Ns . This method has the disadvantage
of requiring the additional storage of 9(Ni )2 numbers in double format, which in general
is greater than the size of the compliance matrix: for large meshes, this method may
become unpractical. Furthermore, this method is slightly more complex to implement
whereas the method proposed in the previous section only requires to solve equation
KU = R with a sparse matrix K. However, the condensation method is well-suited when
the rigidity matrix is very ill-conditioned (Poisson ratio very close to 0.5) in which case
the preconditioned conjugated gradient algorithm may fail.
The second algorithm for computing the compliance matrix Gss is to use the Bound-
ary Element Method (BEM) (C ANAS and PARIS [1997]) instead of the Finite Element
Method (FEM). The algorithm proposed by JAMES and PAI [1999] creates the stiffness
matrix K ss directly from the triangulated surface of the object.
The differences between BEM and FEM are well-understood (H UNTER and P UL -
LAN [1997]). The main advantage of BEM techniques is that they do not require a
volumetric tetrahedral mesh but only its triangulated surface. While there exist several
free software6 for automatically creating tetrahedral meshes from triangulated surfaces
(S IMAIL, OWEN [2000], J OE [1991]), having a control over the final number of vertices
and the quality of tetrahedral elements is still an issue.
On the other hand, BEM techniques have several disadvantages over FEM. First, they
make strong hypotheses about the nature of the elastic material: only homogeneous and
isotropic linear elastic materials can be modeled. Second, the computation of the com-
pliance matrix, and above all its diagonal elements, is difficult to implement and often
numerically unstable because singular integrals must be evaluated over each triangle.
The quality of the triangle geometry can influence the stability of this computation.
Third, this method cannot compute the displacement of any interior point, which can
be a limitation when the displacement of internal structures (vessels, tumors, . . .) is
needed. Finally, the BEM presented in JAMES and PAI [1999] uses centroid collocation
to compute the stiffness matrix. Thus, this matrix allows to compute the displacements
of the centroids of all triangles but not the displacements of the triangulation vertices.
Therefore, the mesh being deformed is not the original triangulated mesh but its dual
mesh which is called a simplex mesh (D ELINGETTE [1999]). Mapping the displace-
ments of triangle centroids into the displacements of vertices is not trivial since the
duality between triangulation and simplex meshes is not a one-to-one mapping from the
geometrical standpoint (D ELINGETTE [1999]).
To conclude, the algorithm proposed by James et al. is more difficult to implement
than our method and it is only suitable for simple material. However, when there is
no software program for creating tetrahedral meshes from triangulations, this approach
should be used.
5.4. On-line computation
5.4.1. Data structure

Before starting the simulation, the compliance matrix Gss , previously stored into a file
as described in Section 5.3.1, is loaded into a specific data structure. Indeed, this data
6 A list of available software can be found at the following two URLs: http://www-users.informatik.
rwth-aachen.de/~roberts/meshgeneration.html and http://www.andrew.cmu.edu/user/sowen/softsurv.html.
structure only describes the triangulated surface shell of the volumetric tetrahedral mesh
with a list of surface vertices and a list of surface triangles. Note that the number of
surface vertices is usually greater than Ns because some surface vertices have an im-
posed displacement. For display purposes, the triangulated data structure may contain
additional information such as 2D or 3D texture coordinates as well as parameters de-
scribing the rendered material. Finally, the data structure contains a list of imposed
displacements and applied nodal forces as a storage of boundary conditions.
ji
For each free vertex of index i, an array of 3 × 3 matrices [Gss ], for all j ∈
ji
{0, . . . , Ns − 1}, is stored inside the vertex data structure. These Ns matrices [Gss ] allow
to compute the displacement of all surface vertex j , once a force is applied on vertex i.
The data structure optimizes the computation time of deformation but at the cost of
being less efficient in terms of memory requirement. Indeed, the compliance matrix Gss
is a symmetric matrix, but it is stored as a non-symmetric matrix in this data structure.
To optimize memory at a small additional computational cost, one could alternatively
ji
store the symmetric matrix as a double array of 3 × 3 compliance matrices [Gss ] which
is filled only if i < j .
5.4.2. Algorithm description and collision processing

The sketch of the algorithm is given in Algorithm 2 and includes two independent parts.
The first part, between lines 1 and 8, consists in detecting and computing the contact be-
tween the soft tissue model and each virtual surgical instrument. In Fig. 5.2, we present
an example of contact between a liver model and a tool. The collision detection algo-
rithm (L OMBARDO , C ANI and N EYRET [1999]) makes the assumption that the handle
and the tool extremity can be approximated by a set of cylinders with rectangular sec-
tion. Its efficiency depends on the availability of graphics cards since it relies on the
1: Reset the list of imposed displacement ldisplacement to the empty list

2: Reset the list of applied forces lforce to the empty list
3: Reset the position of free surface vertices to their rest position + U0ss
4: for all Surface Tools ST i do
5: if collision between the soft tissue model and ST i then
6: Add imposed displacement to the list ldisplacement
7: end if
8: end for
9: if ldisplacement is not empty then
10: Compute the list of applied forces lforce from ldisplacement
11: for all Applied forces F j on vertex j in lforce do
12: for all Free surface vertex k do
kj
13: Add to current position of vertex k, the displacement [Gss ]F j
14: end for
15: end for
16: end if
A LGORITHM 2. On-line computation of mesh deformation.

F IG . 5.2. Example of collision computation between the handle (a) and the extremity (b) of a surgical tool
and a liver soft tissue model (P ICINBONO , L OMBARDO , D ELINGETTE and AYACHE [2002]). The position of
triangles displayed in light gray have been displaced such that the tool is tangent to the liver surface.
OpenGL (W OO , N EIDER and DAVIS [1997]) library. Once a collision has been de-
tected, the collided triangles must be moved such that the tissue model is no longer in
contact with the surgical tool. This computation turns out to be quite complex since it
not only depends on the tool position but also on its trajectory. The algorithm is de-
scribed in P ICINBONO , L OMBARDO , D ELINGETTE and AYACHE [2002]. The outcome
of this computation is a list ldisplacement of imposed displacements that should apply on
each vertex of the collided triangles.
5.4.3. Imposing displacements

The second part of Algorithm 2, between lines 9 and 16 computes the position of all
surface vertices, given the list of imposed displacements.
The first task corresponding to line 10 consists in computing the set of forces {F j }
that should be applied to each vertex j of ldisplacement in order to bring the displacement
of these vertices to Ubj .
To be more didactic, we first consider that only one vertex displacement Ubj is im-
posed on a vertex of index j . Without any collision with a surgical tool, this vertex
has a displacement U0j under the application of the normal boundary conditions (grav-
ity forces, pressure forces, . . . described in Section 4.9). Because the material is linear
F IG . 5.3. Principle of superposition when applying two forces F i and F j to the two nodes i and j .
elastic, it follows the superposition principle: the displacements resulting from the ap-
plication of two sets of nodal forces is the sum of the displacements resulting from the
application of each set of forces. Thus, the force F j to be computed is the force that
should be applied on vertex j in order to create a displacement of that vertex equal to
jj
Ubj − U0j . Because the quantity [Gss ]F j gives the additional displacement of vertex j
resulting from the application of force F j , the force F j is given by
jj −1 b
F j = Gss Uj − U0j .
When the displacements of two vertices i and j are imposed, the problem is slightly
more complex. Indeed, the application of force F i on vertex i not only displaces vertex i
ij
of the amount [Gii ss ]Fi , but it also moves vertex j by the amount [Gss ]Fi (see Fig. 5.3).

ij
Since Fj also displaces vertex i of [Gss ]Fj , to compute the applied force, a 6 × 6

symmetric linear system of equations needs to be solved,

 ij
 Giiss Fi + Gss Fj = Ui − Ui ,
b 0
 Gjssi F + Gjj
ss Fj = Uj − Uj .
b 0
i
Similarly, when the list of imposed displacements ldisplacement contains p elements, then
a symmetric linear system of equations of size 3p × 3p needs to be solved to find the
set of nodal forces. If we use the set of indices ij , j ∈ [1, . . . , p] to denote the set of
vertices where a displacement Uij is imposed, then this linear system of equations can
be written as
 i1 ,i1 i1 ,i2 i ,i     b 
Gss Gss · · · Gss1 p Fi1 Ui1 − U0i1
  .  
 Gi2 ,i1 Gi2 ,i2 · · · ..  .   .. 

 ss ss .  .   . .
  .  =  .  (5.5)
 .
. .
. . . .
.   .
.   .
. 
 . . . . 
ip ,i1 i ,i F ip Ubip − U0ip
Gss ··· · · · Gssp p
In Fig. 5.4, we show an example of a mesh where the same displacement is im-
posed on three vertices. In this particular case, the direction of computed forces departs
strongly from the direction of the prescribed displacement.
F IG . 5.4. (Right) The same displacement is imposed on the three vertices of a triangle; (left) from Eq. (5.5)
we compute, the three forces that should be applied on these three vertices to move them of the given dis-
placement.
5.4.4. Results
Once the set of nodal forces is computed, the additional displacement on all surface (and
potentially internal) nodes are computed as described in lines 11 to 15 of Algorithm 2.
The number of matrix–vector operations is p × Ns for p applied forces. In general,
p, the number of vertices collided with the surgical tools, is small (from 3 to 20) when
compared to Ns (see Fig. 5.5). This is why we chose to store the Ns array of compliance
ji
matrix [Gss ] at vertex j , in order to optimize the inner loop (lines 12 to 14).
The computational efficiency of this quasi-static precomputed model on the liver
mesh shown in Fig. 5.1 is presented in Table 5.1. These performances, measured on
three different hardware platforms, correspond to the frequency update that can be
achieved when running Algorithm 2 in a loop without any computation for visual and
haptic rendering.
F IG . 5.5. Liver deformation based on a linear elastic pre-computed model (C OTIN , D ELINGETTE and AY-
ACHE [1999b]). Solid lines indicate the imposed displacements.
TABLE 5.1
Computation efficiency of quasi-static precomputed linear elastic model for dif-
ferent boundary conditions: either when applying nodal forces or when imposing
displacements
Simulation frequency Pentium PIII

(liver model with 614 surface nodes) 600 MHz
Force applied on 1 node 3772 Hz

Force applied on 5 nodes 754 Hz
Imposed displacements on 1 node 3759 Hz
5 nodes 561 Hz
10 nodes 185 Hz
20 nodes 40 Hz
When applying one nodal force, corresponding to the execution of lines 12 to 14

in Algorithm 2, the computation time is nearly equal to 0.3 ms. The time required to
compute the mesh deformation when applying p forces is strictly proportional to this
value: 0.3 × p ms.
When imposing p displacements, which is what occurs in practice in a surgical sim-
ulator, the additional computation is the solution of a 3p × 3p linear symmetric system
of equations. For p = 1, the overhead is very small and hardly perturbs the simulation
frequency. However, for larger value of p, the overhead becomes dominant. For 20 ver-
tices, for instance, solving the system of equations of size 60 × 60 is 3 times more costly
than computing the 20 ∗ 614 = 1280 matrix–vector products and additions.
5.4.5. Discussion
As a whole, the proposed method is “very efficient”, since it allows real-time visual ren-
dering even for large meshes. When the material is soft enough and when the number
of collided vertices remains small (typically less than 15), this model can also be com-
patible with real-time haptic rendering. In fact, it is one of the few algorithms which are
suitable for the first software architecture described in Section 1.3.2 (see also Fig. 1.6(a))
consisting of one synchronous loop including visual and haptic rendering. Furthermore,
our approach has one major advantage for haptic rendering computation: it already pro-
vides the nodal reaction forces through the algorithm described in Section 5.4.3. Indeed,
the set of forces F ij corresponds to the set of physical forces that have been applied on
each node of index ij in order to deform the soft tissue model: thus, −F ij corresponds
to the nodal reaction force. From this set of forces, one can easily compute the reaction
force along the direction of the tool, as well as the torque at the extremity of the tool.
Using the terminology introduced in Section 1.3.2, we can also state that the quasi-
static precomputed linear elastic model has a very low relaxation time (or equivalently
that it has a high speed of convergence). Indeed, each time Algorithm 2 is run, the soft
tissue is deformed to its static equilibrium position. Because this algorithm can be run
at a high frequency, as seen in Table 5.1, this implies that the relaxation time is very
1: for all Free surface vertex k do

2: if k ∈
/ ldisplacement then
3: Let pk be the position of vertex k after Algorithm 2
previous
4: Let pk be the position of vertex k at the previous iteration.
previous
5: pk ⇐ γ pk + (1 − γ )pk
6: end if
previous
7: pk ⇐ pk
8: end for
A LGORITHM 3. Additional part of Algorithm 2 that adds a visco-elastic behavior controlled by delay para-
meter γ .
low. In fact, for some soft tissue, this time is too low and degrades the visual realism
of the simulation. This is the case, for instance, when the operator grasps and displaces
some soft tissue and suddenly ceases the grasping. Because the model has no longer
any displacements imposed on its surface, it returns in one iteration to its rest position,
while in reality, it takes several milliseconds.
To add some visco-elastic behavior, one can increase the relaxation time artificially
by using a delay function. This approach is described in Algorithm 3. For vertices which
are not colliding with a surgical tool, the final vertex position is a weighted sum between
the position computed by Algorithm 2 and the vertex position at the previous iteration.
The weight parameter 0 γ 1 controls the damping of the material deformation: for
γ = 1, the deformation is not damped (quasi-static motion) while for γ = 0, the motion
is infinitely damped (no motion). Any intermediate value of γ modifies the relaxation
time of the material. Note that this damping is not applied to vertices colliding with
tools because the collision would otherwise appear visually unrealistic. Algorithm 3
assumes that the model has a damping matrix C which is proportional to the identity
matrix: more sophisticated hypotheses (but often more computationally intensive) could
be proposed.
6. Dynamic linear elastic model
In this section, we describe two different soft tissue models that are able to address
with the limitation of the previous model: the simulation of tissue cutting. Using the
terminology defined in Section 1.3.2, these two methods can be qualified as “Explicit
Iterative Methods” sharing the advantage of requiring a small computation time for each
iteration but with the drawback of having a low speed of convergence.
The main difference between these two models is that the first can model the visco-
elastic behavior of the soft tissue properly whereas the second does not require the
evaluation of any time step and is unconditionally stable.
Finally, we propose in Section 6.3 a hybrid model which combines any of the two
previous models with the precomputed linear elastic model seen in Section 5.
6.1. Tensor–mass model
6.1.1. Introduction
The tensor–mass model is based on the dynamic law of motion described in Eq. (4.22):
MÜ + CU̇ + KU = R.
This second order differential equation couples the motion of tissue under the influence
of inertia MÜ, of visco-elasticity CU̇, elasticity KU and external loads R.
The most efficient way to solve the equation above is by far to use modal analysis
(BATHE [1982]). By making simple assumptions about the damping matrix C, it is pos-
sible to simplify the above PDE into a small set of ordinary differential equations with
an appropriate change of basis. The proper basis is given by the eigenvectors associated
to the largest eigenvalues of the generalized eigenproblem Kφ = ω2 Mφ.
However, the eigenproblem must be solved each time the rigidity matrix is modified.
Therefore, this approach is not suitable for simulating tissue cutting, since the compu-
tation cost to solve the eigenproblem is very high.
Instead, a classical method to solve Eq. (4.22), is to use integration methods: the time
dimension is uniformly discretized with a time step t, and each term of that equation is
supposed to be constant during each time interval. There is an important distinction be-
tween implicit integration schemes and explicit integration schemes depending whether
the position of the model at time t + t requires the solution or not of a global linear
systems of equations (see also the discussion in Section 1.3.2).
Implicit schemes are unconditionally stable which allows the use of large time steps.
In structural analysis, the Houbolt method (H OUBOLT [1950], BATHE [1982]) and the
Newmark method (N EWMARK [1959], BATHE [1982]) are the most commonly used.
However, these schemes require either to inverse a sparse matrix or to solve at each
iteration a linear system of equations. Considering the time required to solve such a
linear system (a few seconds for a small-size mesh), these implicit schemes cannot be
used for real-time interaction.
Instead, we chose to use explicit integration schemes which have several nice prop-
erties (ease of implementation, low computational cost) compared to implicit schemes
but with the drawback of being conditionally stable: the time step must be smaller than
a critical time step tcritical . Therefore, smaller time step t must be used for explicit
schemes which yields a larger relaxation time and a longer time for convergence.
6.1.2. Mass matrix

Regarding the mass matrix, a common choice consists in replacing the symmetric posi-
tive definite matrix M with a diagonal matrix, where each diagonal element is the sum
of all row elements in the original matrix: this lumped mass matrix is detailed in Sec-
tion 4.8.
In order to keep the time step t large enough during the simulation, we propose
a further simplification of the mass matrix M by considering that the nodal mass is
constant for all nodes, which makes M proportional to the identity matrix,
M = m0 I3 ,
where m0 is the average mass per node computed as the total mass of the tissue divided
by the number of nodes in the initial mesh.
Indeed, the critical time step t of the iterative scheme is inversely proportional to
the highest eigenvalue of the matrix M−1 K, while the speed of convergence is related to
the ratio between the largest to the smallest eigenvalues of the same matrix, also called
the condition number of that matrix.
From the equation of the nodal stiffness matrix [Ki,i ], we can state that the nodal stiff-
ness is proportional to the size (for instance, the largest foot height) of all the tetrahedra
surrounding each node:

1
[Ki,i ] = (λT + µT )(mi ⊗ mi ) + µT A2i I3 .
36V (T )
T ∈S (i)
Thus, the largest eigenvalue of K is determined by the largest tetrahedra while the con-
dition number is given by the size ratio between the largest and smallest tetrahedra.
On the other hand, when performing mass lumping, as in B RO -N IELSEN [1998], the
nodal mass of M−1 is inversely proportional to the volume of tetrahedra surrounding
each node. Therefore, the power spectrum of M−1 K largely differs from that of K: the
largest eigenvalue of M−1 K now becomes related to the tetrahedron of smallest size
while the condition number is related to the square ratio between the largest and small-
est tetrahedra. These properties of M−1 K have two consequences for the simulation of
tissue cutting: both the speed of convergence and the time step t decrease as tetrahedra
of small size are created.
By choosing a mass matrix proportional to the identity matrix, we keep the spectral
properties of the rigidity matrix: the creation of small tetrahedra does not entail any
decrease of the time step and limits the decrease of the speed of convergence. How-
ever, this choice is a gross approximation of physics since the total mass of the tissue
increases as the number of elements increases. As claimed in Section 3.1.5, we prefer
to satisfy real-time constraints of the simulation (by keeping a large value of t) at the
expense of coarse approximations of the tissue dynamic behavior.
6.1.3. Numerical integration

Several explicit iterative schemes can be proposed from Eq. (4.22) depending on the
choice of damping matrix and discretization of time derivatives. Below, we propose
three explicit schemes that are of interest in the context of surgery simulation. In the
remainder, we write t U the displacement vector at time t.
Euler method. This method uses central finite differences to estimate acceleration but
right finite difference to estimate speed. Furthermore, sophisticated damping matrix
such as Rayleigh damping can be employed in this scheme:
m0 t −t 1
2
U − 2t U + t +t U + (γ1 m0 I3 + γ2 K) t U − t −t U + Kt U = t R.
t t
The displacement at time t + t can be computed through the recurrent equation:
t +t

U = t U + (1 − tγ1 ) t U − t −t U

t 2 t γ2 t t t 2 t
−K U+ U − t −t U + R.
m0 m0 m0
Euler method with central finite difference. In this case, central finite differences are
used to estimate both acceleration and speed, while constant damping is used γ2 = 0:
m0 t −t γ1 m0 t +t
2
U − 2t U + t +t U + U − t −t U + Kt U = t R,
t 2t
which leads to the following update equation:
t +t 2 − γ1 t t 2t 2 t
U = tU + U − t −t U − K U − tR . (6.1)
2 + γ1 t m0 (2 + γ1 t)
Runge–Kutta method of order 4. The Runge–Kutta method (P RESS , F LANNERY,

T EUKOLSKY and V ETTERLING [1991]) is an integration method of fourth order of ac-
curacy, but which requires four evaluations of the Euler recurrent equation. To describe
this method, it is necessary to write the original equation as a first order differential
equation,
! ! ! ! R !
d U̇ Ü − mC0 − mK0 U̇
= = + m0 .
dt U U̇ 1 0 U 0
Now, the state of a soft tissue model at time t is described by two vectors: displacement
vector t U and the velocity vector t U̇. Applying the simple Euler method on this system
gives the following relation:
! ! ! !
m0 −C U̇ − K U + R
t +t U̇ t U̇ 1 t t t
δv t U, t U̇
t +t U − tU = t = .
t U̇ δu t U, t U̇
The fourth order Runge–Kutta method requires to compute the following eight incre-
mental displacement and velocity vectors:

δv1 = δv t U, t U̇ , δu1 = δu t U, t U̇ ,

δu1 t δv1 δu1 t δv1
δv2 = δv U +
t
, U̇ + , δu2 = δu U +
t
, U̇ + ,
2 2 2 2

δv3 = δv t U + , U̇ + , δu3 = δu t U + , U̇ + ,
2 2 2 2

δv4 = δv t U + , U̇ + , δu4 = δu t U + , U̇ + .
2 2 2 2
Finally, the velocity and displacement for the next time step are given by the following
equation:
t +t U̇
! t U̇
! ! ! ! !
1 δv1 1 δv2 1 δv3 1 δv4
t +t U = tU + + + + .
6 δu1 3 δu2 3 δu3 6 δu4
TABLE 6.1
Comparison between three explicit integration methods for soft tissue modeling
Euler Euler central Runge–Kutta

method finite differences method
Computation time low low high

Damping Rayleigh basic basic
Time step small medium high
Comparison between the three methods. We summarized in Table 6.1 the properties of
the three methods described above. Three qualitative criteria were proposed to outline
the advantages and drawbacks of each method. In terms of computation time required
to update the position of a model, the first two Euler methods are equivalent while the
Runge–Kutta method is at least four times slower. As far as damping is concerned, only
the first Euler method allows to use Rayleigh damping while the two other methods can
only use diagonal damping matrices. Having a non-diagonal damping matrix helps in
keeping a continuous field of velocity throughout the model which improves the visual
realism of the simulation. Finally, the Runge–Kutta method is more stable than the
Euler method and our experience showed that a tenfold increase of the time step can
be observed in the former case. The Euler method with central finite differences allows
larger time steps than the Euler method because the velocity computation leaps over
position computation by one time step.
6.1.4. Data structure

With explicit schemes, the update of the mesh position can be performed locally, at
the vertex level, without creating any global matrix. Indeed, for each free vertex of
index i, we can take advantage of the sparse nature of the rigidity matrix K, in order to
compute the matrix–vector product KU. More precisely, from Eq. (4.13), it is clear that
the off-diagonal stiffness matrices [Ki,j ] are non-null matrices only when there is an
edge connecting vertices i and j in the tetrahedral mesh. Therefore, only the set N (i)
of vertices connected to vertex i by an edge is involved when computing the elastic
force Fi applied on vertex i. For instance, the update Eq. (6.1) can be computed for a
vertex i as
t +t 2 − γ1 t t
ui = t ui + ui − t −t ui
2 + γ1 t

2t 2
− [Ki,j ]t uj + [Ki,i ]t ui − t Ri .
m0 (2 + γ1 t)
j ∈N (i)
The data structure that is suitable for performing this computation follows the data struc-
ture required for storing a tetrahedral mesh. The basic structure consists in a double-
linked list of vertices, edges and tetrahedra. For each vertex, we store its current posi-
tion t qi , its rest position pi and the symmetric tensor [Ki,i ]. For each edge, we store its
two adjacent vertices (vertex i and vertex j ) as well as the tensor [Ki,j ], as sketched in
Fig. 6.1. We therefore take advantage of the symmetric nature of the stiffness matrix by
storing the off-diagonal rigidity matrix only once.
F IG . 6.1. Representation of the data structure of a tensor–mass model. The 3 × 3 rigidity matrices are stored
at each edge and each vertex. The symmetry of the rigidity matrix enables to store only one tensor per edge.
Finally, for each tetrahedron, we store its four vertices and its six edges as well as
the Lamé coefficients λi , µi , the area vectors mi and if required the direction of aniso-
tropy a0 .
6.1.5. Cutting and refinement algorithms

One of the basic tasks in surgery simulation consists in cutting and tearing soft tissue.
With the dynamic linear elastic model, these tasks can be achieved efficiently.
To perform an hepatectomy (partial resection of the liver), the use of scalpel instru-
ments is not appropriate because of the important vascularization of the liver. Instead,
surgeons usually proceed with a set of pliers that smash hepatic cells or with a cavitron
device that destroys the hepatic parenchyma with ultrasound energy: in both cases, the
resection is performed by removing soft tissue. It is therefore important to simulate the
removal of bits of soft tissue located at the vicinity of a surgical tool. To perform this
simulation, two basic meshing techniques must be implemented: removal of tetrahedra
and local refinement.
At first sight, removing a single tetrahedron from a tetrahedral mesh is straightfor-
ward. However, in order to obtain a visually realistic simulation, one should avoid to
produce isolated or self-intersecting tetrahedra or even tetrahedra connected through
a single vertex. A proper way to keep “visually appealing” meshes is to constrain the
mesh to be a manifold mesh in addition to being a conformal mesh. Indeed, in a manifold
mesh, the shell of a vertex located on the mesh surface is homeomorphic a half-sphere
(the shell is a sphere for interior vertices) which allows to define unambiguously a sur-
face normal at that vertex. However, by adding this topological constraint, even remov-
ing a single tetrahedron is not straightforward as discussed in F OREST, D ELINGETTE
and AYACHE [2002b]. The detailed description of the topological issues relevant to
the operation of tetrahedron removal falls outside the scope of this chapter; instead we
present briefly the algorithms related to the computation of soft tissue deformation.
Once a collision between a surgical tool and a set of tetrahedra has been detected,
each tetrahedron of the set is removed one after the other. After updating the topological
F IG . 6.2. To remove the tetrahedron whose external triangle has been selected (dark gray), it is necessary to
update the local rigidity matrices stored at the vertices and edges of that tetrahedron.
structure of the mesh, the local vertex and edge stiffness matrices must also be updated
(see Fig. 6.2). When removing tetrahedron T , its 6 edge tensors [BTi,j ] and 4 vertex
tensors [BTi,i ] are computed based on Eqs. (4.8) and (4.9) and are subtracted from the
current edge and vertex local rigidity matrices:

[Ki,i ] = [Ki,i ] − BTii , [Ki,j ] = [Ki,j ] − BTij .
These ten local operations are performed efficiently because of the specific data struc-
ture associated with a tetrahedron.
The second meshing technique, local refinement, can be used in two cases. First, it
can be used offline (before the simulation), to increase the mesh resolution at places of
high curvature or near structures of interest (tumors, gall blader, . . . ). Second, it is often
necessary to refine the mesh locally during the removal of soft tissue when the tetrahedra
to be removed are too large. In the former case, sophisticated meshing techniques can
be employed while in the latter case, real-time constraints allow the application of only
basic refinement algorithms. An example of such a basic algorithm consists in adding a
vertex at the middle of an edge and then splitting all tetrahedra adjacent to that edge into
two tetrahedra (see Fig. 6.3). In this case, the edge and vertex tensors of all tetrahedra
adjacent to that edge are first removed and the contributions from all newly created
tetrahedra are then added. A more sophisticated refinement algorithm can be found
in F OREST, D ELINGETTE and AYACHE [2002b].
F IG . 6.3. Local refinement of a tetrahedral mesh. An edge is split into two edges by inserting a vertex. The
rigidity matrices must be updated for vertices and edges that already existed (drawn in dark grey) while these
matrices must be computed for newly created vertices and edges (drawn in light grey).
F IG . 6.4. Deformation of a cylinder subject to gravity forces: some tetrahedra are progressively being re-
moved at its center leading to a separation into independent solids.
The proper adjustment of stiffness matrices during the removal of soft tissue rein-
forces the visual realism of the simulation significantly: this is especially the case when
the tissue is cut while being stretched. For instance, in Fig. 6.4, we show the deformation
of a cylinder being cut: the cylinder is fixed at its upper part and is under the influence
of gravity forces along its main axis.
6.1.6. Algorithm description

Before describing the deformation algorithm for a tensor–mass model, we shortly de-
scribe the initialization stages in Algorithm 4. Once the vertex and edge stiffness matri-
ces have been assembled, it is necessary to estimate a time step t that allow the stabil-
ity of the iterative schemes described in Section 6.1.3. Finding the critical time step (i.e.,
the highest possible time step) is actually a difficult task because of the lack of a closed-
form expression. However, a practical approach is to estimate the critical time step as
a product of an unknown constant with the time step given by the Courant–Friedrich–
Levy condition (P RESS , F LANNERY, T EUKOLSKY and V ETTERLING [1992]):
%
ρ
(t)Courant = lmax .
λ + 2µ
Algorithm 5 presents the different loops required to update a tensor–mass model.
Unlike the precomputed quasi-static model, it is not necessary to maintain an explicit
list of vertices that are displaced by the collision with a surgical tool: it is sufficient
1: for all Tetrahedron T do

2: Compute the 4 area vectors mi
3: for all Vertex i of T do
4: Compute the local rigidity matrix [BTii ]
5: [Ki,i ] ⇐ [Ki,i ] + [BT
ii ]
6: end for
7: for all Edge between vertices i and j of T do
8: Compute the local rigidity matrix [BTij ]
9: [Ki,j ] ⇐ [Ki,j ] + [BT
ij ]
10: end for
11: end for
12: Estimate time step t.
A LGORITHM 4. Matrix assembly for the tensor–mass model performed before any simulation.
1: for all Surface tools ST i do

3: if ST i represent a cavitron device then
4: Eventually refine locally the mesh near the collision
5: Remove tetrahedra located near the extremity of ST i
6: end if
7: Impose displacements on vertices near the contact zone and raise a flag
on these vertices
8: end if
9: end for
10: for all edge e connecting vertex i and j do
11: add elastic force [Ki,j ]t ui to vertex i
12: elastic force [Ki,j ]Tt uj to vertex j
13: end for
14: for all vertex i do
15: if vertex i is free (flag not raised) then
16: compute elastic force [Ki,i ]t ui
17: update vertex position t pi based on one of the three iterative schemes
described in Section 6.1.3
18: else
19: reset flag
20: end if
21: end for
A LGORITHM 5. On-line computation of tensor–mass model.
(see line 7) to raise a flag stating that these vertices are not free vertices. A second
important feature of this algorithm is the existence
of a loop on the mesh edges in
order to compute the matrix–vector products j ∈N (i) [Ki,j ]t uj . This approach is more
efficient than scanning iteratively the neighbors N (i) for each vertex i. When using the
fourth order Runge–Kutta algorithm, the algorithm from lines 10 to 21 must be modified
since it is then necessary to scan four times the edges and vertices of the mesh. For the
Euler method, only lines 11 and 12 must be modified in order to compute
2
t t γ2 t t t −t

K U+ U− U
m0 m0
instead of Kt U.
6.1.7. Comparison between spring–mass and tensor–mass models

We have used the word “tensor–mass model” to designate a finite element model based
on Newtonian dynamics and discretized with an explicit scheme. This word has been
chosen in order to stress the similarity between a “tensor–mass model” and a “spring–
mass model”. In particular, it is the purpose of this section to oppose to the widely spread
belief stating that “finite element models are slower and more complex to implement
than spring–mass models”.
A spring–mass model (BARAFF and W ITKIN [1998]) consists of a set of masses and a
set of springs connecting these masses. The force applied to a point pi in a spring–mass
system, is given by the relation

pi pj
Fi = kij pi pj − lij0 , (6.2)
pi pj
j ∈N (i)
where kij is the stiffness coefficient between vertices i and j , lij0 is the length at rest.
Similarly, on a tensor–mass model, the elastic force applied on vertex i is given by

Fi = [Ki,i ]ui + [Ki,j ]uj . (6.3)

j ∈N (i)
By comparing Eqs. (6.2) and (6.3), it is clear that both dynamic models have the same
computational complexity which is linear in the number of edges. In practice, we have
observed a slight computational advantage for the tensor–mass model, mostly because
it does not include any square root evaluation.
However, both approaches differ substantially in terms of biomechanical modeling.
Spring–mass systems constitute a discrete representation of an object and their behavior
strongly depends on the topology of the spring network. Adding or removing a spring
may change the elastic behavior of the whole system drastically. Conversely, a finite
element model is a continuous representation of the object and its behavior is inde-
pendent of the mesh topology (it mostly depends on the mesh resolution). This is an
advantage when mesh cutting is performed since it produces continuous and natural
deformations.
Because all biomechanical data related to biological soft tissue are formulated as
parameters found in continuum mechanics (such as Young’s modulus or Poisson coef-
ficients), it is a priori difficult to model realistic soft tissue deformations with a spring–
mass system. However, several authors (L OUCHET, P ROVOT and C ROCHEMORE
[1995], D EUSSEN , KOBBELT and T UCKE [1995]) have developed genetic or simulated
TABLE 6.2
Comparison between the three soft tissue models: pre-computed quasi-static,
tensor–mass and spring–mass models
Pre-computed Tensor–mass Spring–mass
Computational efficiency +++ + +

Biomechanical realism + + −
Cutting simulation − ++ +
Large displacements − − +
annealing algorithms to identify spring parameters (stiffness and damping) from a set
of known deformations of an object.
Finally, as previously mentioned, the tensor–mass model is only valid for small dis-
placements. This model is invariant under the application of a global translation, but
if a global rotation is applied to the rest shape Mrest , then the forces applied to all
vertices will not be null. On the opposite, a spring–mass model under the same dis-
placement would not deform, since the length of the springs are preserved under a rigid
transformation. The difference between these three soft tissue models is summarized in
Table 6.2.
6.2. Relaxation-based elastic models
6.2.1. Introduction
In this section, we introduce an alternative algorithm to the tensor–mass model. This
algorithm is based on Gauss–Seidel relaxation and has the following properties:
• Its iterative scheme is unconditionally stable. It does not require the estimation of
any critical time step.
• The relaxation algorithm is fairly efficient (small computation time required for
one iteration) but it is slightly less efficient than a tensor–mass model.
• The algorithm is based on static equilibrium equations whereas tensor–mass mod-
els are based on the dynamic law of motion.
• The position of each vertex is updated asynchronously, one vertex after the other.
However, when compared to tensor–mass models, relaxation-based elastic models have
two drawbacks. First, their implementation requires the following property for the mesh
data structure: each vertex should be able to access efficiently its adjacent edges. This
topological “vertex–edge” relationship can be stored in two ways inside a data structure.
In a first approach, a list of edges can be stored explicitly at each vertex. After removing
or adding tetrahedra, the edge list must be updated for all vertices belonging to these
tetrahedra. To achieve this update, each edge must have a list of adjacent tetrahedra
which should also be explicitly updated upon the removal or addition of tetrahedra.
In a second approach, the list of edges adjacent to a vertex is recovered through
the knowledge of a single tetrahedron adjacent to this vertex. This approach is only
applicable if we constrain the tetrahedral mesh to be a manifold mesh (see F OREST,
D ELINGETTE and AYACHE [2002b] for more details). Indeed, in such case, the neigh-
borhood of a vertex is homeomorphic to a topological sphere or half-sphere. By march-
ing around a vertex from a given tetrahedron, it is possible to obtain all tetrahedra ad-
jacent to a given vertex and consequently the list of all adjacent edges. In this case as
in the former case, we do store a list of adjacent edges for each vertex in order to avoid
duplicating the search algorithm. However, when a tetrahedron is removed or added,
this topological list is reseted and the pointer to the adjacent tetrahedron is eventually
updated.
The second drawback of relaxation algorithms is that they require in average 3 times
more storage than the tensor–mass model. Indeed, in addition to the symmetric stiffness
matrix, a non-symmetric stiffness matrix must be stored.
6.2.2. Overview of the algorithm

Following the notations of Eq. (6.3) the static problem KU = R can be written at the
level of each vertex i as

[Ki,i ]ui + [Ki,j ]uj = Ri . (6.4)

j ∈N (i)
For relaxation algorithms, the displacement of a vertex ui is updated independently
from other vertices. Therefore, the notation t +t ui to describe the position of vertex i
at the next time step cannot be used, since formally there is no temporal evolution (and
no temporal variable t) in relaxation algorithms. Thus, we note + ui the next position of
vertex i and ui its current position.
The principle of relaxation algorithms is quite straightforward: each vertex is moved
in order to locally solve Eq. (6.4). Thus, the displacement + ui is given by

+
ui = − [Ki,i ]−1 [Ki,j ]uj + [Ki,i ]−1 Ri . (6.5)
j ∈N (i)
This is equivalent to minimizing the total mechanical energy by successively optimiz-
ing each variable ui . It is therefore similar to the Iterative Conditional Mode (ICM)
algorithm (B ESAG [1986]) used in statistical analysis.
If all displacements {ui } are successively updated according to Eq. (6.4), then this
method is equivalent to the Gauss–Seidel relaxation method (S AAD [1996]). More pre-
cisely, we can decompose the stiffness matrix K as the sum of three terms: KD a 3 × 3
block diagonal matrix, KC the lower triangle matrix of K and KTC the upper triangle
matrix of K:
 [K ] ··· 0   0 
1,1 0 0 ··· 0
   ..
 0
.
[K2,2 ] . . 0   [K2,1] 0 . 0 
K= .  +  .. ..  +KTC .
 . ..
.
..
.
..   .
..
.
..
. . 
. .
0 0 · · · [KN,N ] [KN,1 ] [KN,2 ] · · · 0

KD KC
With this notation, the Gauss–Seidel relaxation consists in the application of an iterative
equation

k+1
U = (KD + KC )−1 −KTC k U + R , (6.6)
where k U is the displacement vector at iteration k.
To speed-up convergence, we use over-relaxation (known as the Simultaneous Over-

Relaxation algorithm (S AAD [1996])) that consists in anticipating future correction with
an overrelaxation parameter ω,

k+1
U = (KD + ωKC )−1 −ωKTC k U + (1 − ω)KD k U + ωR . (6.7)
This equation translates at the vertex level with the recursion

+
ui = (1 − ω)ui − ω [Ki,i ]−1 [Ki,j ]uj + ω[Ki,i ]−1 Ri . (6.8)
j ∈N (i)
If ω = 1, then the SOR algorithm is equivalent to the Gauss–Seidel relaxation. Con-

vergence is guaranteed for values of ω comprised between 1 and 2, while fastest con-
vergence is obtained for a critical value
2
ωoptimal = √ ,
1+ 1 − ρGS
where ρGS is the spectral radius (the modulus of the largest eigenvalue) of the matrix
(KD + ωKC )−1 KTC .
The overrelaxation parameter ω controls the dynamics of the soft tissue model. With
ω ≡ 2, the model tends to overshoot around the solution whereas with ω ≡ 1, the motion
is very damped. In practise, we chose a value of ω = 1.2 as a trade-off between these
two behaviors.
6.2.3. Algorithm description

The application of the SOR recursive Eq. (6.8) requires the computation of matrices
[Ki,i ]−1 [Ki,j ] and [Ki,i ]−1 . For speed-up purposes, these matrices are stored respec-
tively at each vertex and edge. Because the matrix K−1 D K is no longer symmetric, at
each edge linking vertices i and j , we store the two 3 × 3 matrices [Ki,i ]−1 [Ki,j ] and
[Kj,j ]−1 [Ki,j ]T .
The algorithm of the relaxation-based elastic model is presented as Algorithm 6.
A large part is dedicated to the update of these additional matrices each time a topolog-
ical change of the mesh occurs. A flag is positioned at each vertex and edge in order
to indicate whether matrices [Ki,i ]−1 [Ki,j ] and [Ki,i ]−1 are up-to-date or not. This flag
is raised each time a topological change takes place at a vertex or edge level and it is
lowered once these matrices are updated.
6.3. Hybrid models
6.3.1. Motivation
We have previously described two types of linear elastic models:
(1) a quasi-static pre-computed elastic model which is computationally efficient but
that does not allow any change of topology (cutting, tearing) (see Section 5).
(2) two dynamic elastic models (tensor–mass and relaxation-based models) that have
lower convergence speed but that allow topology changes (see Sections 6.1
1: for all Surface tools STi do

3: if ST i represents a cavitron device then
4: Possibly refine locally the mesh near the collision
5: Remove tetrahedra located near the extremity of ST i
6: end if
7: Impose displacements on vertices near the contact zone
8: end if
9: end for
10: for all free vertex i do
11: if flag raised at vertex i then
12: compute and store [Ki,i ]−1
13: lower flag at vertex i
14: end if
15: u i ⇐ (1 − ω)ui + ω[Ki,i ]−1 Ri
16: for all edge e connecting vertex i and j do
17: if flag raised at edge e then
18: if flag raised at vertex j then
19: compute and store [Kj,j ]−1
20: lower flag at vertex j
21: end if
22: compute and store [Ki,i ]−1 [Ki,j ] and [Kj,j ]−1 [Ki,j ]T
23: lower flag at edge e
24: end if
25: u i ⇐ u i − ω[Ki,i ]−1 [Ki,j ]uj
26: end for
27: ui ⇐ u i
28: end for
A LGORITHM 6. On-line computation of the relaxation-based model.
and 6.2). In the remainder, we use tensor–mass models as the method for de-
forming.
To combine these two approaches, we make a distinction between two types of anatom-
ical structures that usually appear in a surgical simulation:
• Anatomical structures which are the target of the surgical procedure. On these
structures, tearing and cutting need to be simulated. In many cases, they correspond
to pathological structures and only represent a small subset of the anatomy that
needs to be visualized during the simulation.
• Anatomical structures which only need to be visualized or eventually deformed but
which are not submitted to any surgical action.
Thus, in a hybrid model, we propose to model the former type of anatomical structures
as tensor–mass models whereas the latter type of structures should be modeled as a
pre-computed linear model. However, this method is only efficient if the number of
tensor–mass elements is kept as low as possible.
F IG . 6.5. (a) Definition of the interface nodes in a hybrid elastic model; (b) hybrid elastic model with eight
interface nodes (C OTIN , D ELINGETTE and AYACHE [2000]).
6.3.2. Description
A hybrid elastic model Mhybrid is composed of two different types of elements: let
Mdynamic be the set of tensor–mass elements and let Mquasi-static be the set of pre-
computed linear elastic elements. The model Mdynamic is connected to Mquasi-static by
a set of common vertices called interface nodes. These interface nodes define additional
boundary conditions for each model. As seen in Fig. 6.5, the two models may not be
completely connected along their entire boundaries. In fact, a way to reduce the number
of tensor–mass elements, is to associate a fine pre-computed elastic model with a coarse
tensor–mass model. As shown in Fig. 6.5(b), this incomplete interface causes some
visual artifacts due to the non-continuity between two neighboring parts. However, if
the interface zone between the two elastic models is not an important visual cue, a
different mesh resolution can be used.
Since both linear elastic models follow the same physical law, their combination
should behave exactly as a global linear elastic model. Thus, the additional boundary
conditions imposed at the interface nodes must be consistent with responding terms of
forces and displacements for both models.
Fig. 6.6 summarizes the computation loop of a hybrid model. Since the pre-computed
model Mquasi-static is more efficient with force boundary conditions than with imposed
displacements (see Section 5.4.3), its update is based on forces applied at interface
nodes by Mdynamic but also on imposed displacements resulting from the contact with
surgical tools. The applied forces originating from Mdynamic are computed as reaction
forces (opposite of elastic force) at interface nodes. At this stage, the displacement of
all surface nodes of Mquasi-static is computed and the position of interface nodes be-
comes new displacement constraints for Mdynamic . After Mquasi-static , Mdynamic is up-
dated based on displacements imposed at the interface nodes by Mquasi-static and the
displacements imposed by the user interaction.
6.3.3. Examples
In Fig. 6.7, we present an example of a hybrid cylinder model undergoing deformation
caused by gravity forces. The different stages of the deformation process are shown.
F IG . 6.6. Interaction loop for a hybrid elastic model. Both models are updated alternatively while allowing
for user interaction.
F IG . 6.7. Deformation of a hybrid elastic model under a gravity force: the upper cylinder consists of a
pre-computed linear elastic model whereas the lower part is a tensor–mass model. The leftmost figure corre-
sponds to the initial position of the mesh and the rightmost figure to the equilibrium state.
When the equilibrium is reached, as shown in the rightmost figure, forces applied at
the interface nodes are null and displacement vectors stabilize to a constant value. In
this example, both quasi-static and dynamic models have the same elastic properties
and we verified that the equilibrium position is the same as the one that would have
been reached by a single quasi-static or dynamic elastic model. Furthermore, this hybrid
model converges significantly faster than the corresponding dynamic elastic model.
The second example is related to the simulation of hepatectomy, i.e., the removal
of one of the eight anatomical segments – known as Couinaud segments (C OUINAUD
[1957]) – of a liver. In this example the segment number six has to be removed. A tetra-
hedral mesh of a liver has been created from a CT scan image. It is composed of 1537
vertices and 7039 tetrahedra – see Fig. 6.8. The tetrahedra of the sixth anatomical seg-
ment, which represent 18% (280 vertices and 1260 tetrahedra) of the global mesh, are
F IG . 6.8. Display of a hybrid liver model. The part displayed in blue corresponds to the pre-computed
quasi-static elastic model whereas the red part corresponds to the tensor–mass model. The interface nodes
ensure the visual continuity between the two elastic models.
modeled with a tensor–mass model and the remaining tetrahedra with a pre-computed
linear elastic model.
In Fig. 6.9, we show different stages of the hepatectomy simulation. The first six
pictures show the deformation of the model when the tool collides with the dynamic
model. Since both models have the same elastic characteristics, it is not possible to
visually distinguish the interface between the two different elastic models.
The last six pictures show the cutting of the liver segment by removing additional
tetrahedra. The cutting occurs for the tetrahedron being collided by the tool. One can
notice that each part of the hybrid model deforms naturally itself during the resection
simulation.
F IG . 6.9. Different stages of the simulation of hepatectomy. In this simulation, we have included lineic mod-
els of the main bifurcations of the portal vein (F OREST, D ELINGETTE and AYACHE [2002b]). The simulation
consists in removing some hepatic parenchyma but also to clamp and cut each vessel.
7. Large displacement non-linear elastic model
7.1. Shortcomings of linear elasticity
The physical behavior of a soft tissue may be considered as linear elastic for small dis-
placements and small deformations (F UNG [1993], M AUREL , W U , M AGNENAT T HAL -
MANN and T HALMANN [1998]). The hypothesis of small displacements corresponds to
displacements that are typically less than 10% of the mesh size.
In the context of surgery simulation, this hypothesis is often violated. For instance,
the lobes of the liver are often folded to access underlying structures such as the gall
bladder. Also during the resection of a soft tissue, it is common that pieces being cut
undergo large rotations either under the action of gravity or under the action of surgical
instruments.
In such cases, linear elasticity is not an appropriate physical model because it makes
the assumption of infinitesimal strain instead of finite strain. To exhibit the shortcomings
of linear elasticity, we produced two examples pictured in Figs. 7.1 and 7.2.
In a first example, we illustrate the action of a global rotation on a linear elastic model.
When an object (an icosahedron in Fig. 7.1) undergoes a global rotation, its elastic
energy increases, leading to a large variation of volume (as seen in the wireframe mesh
of the rightmost figures). Indeed, the infinitesimal strain tensor EL (X) = 12 (∇U + ∇UT )
is not invariant when a global rotation R is applied since in this case ∇U = R − I3 and
therefore EL (X) = 12 (R+RT )−I3 = [0]. The two invariants (tr EL )2 and tr E2L increases
under rotation as does the elastic energy.
The second example shows the effect of linear elasticity when only one part of an
object undergoes a large rotation (which is the most common case). The cylinder pic-
tured in Fig. 7.2 has its bottom face fixed while a force is being applied at the central top
vertex. The arrows correspond to the trajectories of some vertices: because of the linear
elastic hypothesis, these trajectories are straight lines. This results in unrealistic distor-
tions of the mesh. Moreover, abnormal deformations are not equivalent in all directions
since the object only deforms itself in the rotation plane (Fig. 7.2(c) and (d)).
7.2. St Venant–Kirchhoff elasticity
To overcome the limitations of linear elasticity, we proposed to adopt the St Venant–

Kirchhoff elasticity. The St Venant–Kirchhoff model is a generalization of the linear
F IG . 7.1. Global rotation of the linear elastic model (wireframe).

F IG . 7.2. Successive deformations of a linear elastic cylinder (P ICINBONO , D ELINGETTE and AYACHE
[2001]). (a) and (b): side view; (c) and (d): top view.
model for large displacements, and is a particular case of hyperelastic materials. It has
been used to model various materials (Table 3.8.4 of C IARLET [1987] provides the con-
stants for materials like steel, glass, lead or rubber) including facial soft tissue (G LAD -
ILIN [2002]) and trabecular bone (BAYRAKTAR , A DAMS , G UPTA , PAPADOPOULOS
and K EAVENY [2003]). A St Venant–Kirchhoff material relies on the Hooke’s law as
the definition of elastic energy (see Eq. (3.5) in Section 3.2.3) but the linearized strain
tensor EL is replaced by the Green–Lagrange strain tensor E:
1
E(X) = ∇U + ∇UT + ∇UT ∇U , (7.1)
2
λ
WNL (X) = (tr E)2 + µ tr E2 . (7.2)
2
The Green–Lagrange strain tensor E is no longer a linear function of the displacement
field. A first property is that the elastic energy becomes invariant under the application
of rotations. Indeed, when a rigid transformation (with rotation matrix R) is applied
to an object, the gradient of the displacement field is ∇U = R − I3 and therefore the
Green–Lagrange strain tensor remains zero (since R RT = I3 ),

1
E(X) = R − I3 + RT − I3 + RT − I3 (R − I3 )
2
1
= R + RT − 2I3 + RT R − R − RT + I3
2
= [0].
A second property is that the elastic energy WNL (Section 3.2.3), which was a quadratic
function of ∇U in the linear case, is now a fourth-order polynomial function with respect
to U:
λ
WNL = (tr E)2 + µ tr E2
2
!2
λ 1 µ
= (div U) + ∇U 2
+ µ∇U2 − rot U2
2 2 2
µ& &2
+ µ(∇U : ∇U ∇U) + &∇U ∇U& ,
t t
(7.3)
4
λ λ
WNL = WLinear + (div U)∇U2 + ∇U4
2 8
µ& t &2
+ µ ∇U : ∇U ∇U +t &∇U ∇U& ,
4

where WLinear is given by Eq. (3.5) and A : B = tr(At B) = i,j aij bij is the dot product
of two matrices.
Furthermore, we can extend this isotropic non-linear elastic energy to take into ac-
count “transversally isotropic” materials as performed in Section 3.2.4 for the linear
elastic model. In fact, Eq. (3.9), which defines the additional anisotropic term, still holds
for St Venant–Kirchhoff elasticity. However, for the sake of clarity, we chose to keep
only the anisotropic contribution which penalizes the material stretch in the direction
given by unit vector a0 :

λ 2
WTrans_iso = WNL + − + µ at0 Ea0 ,
2
where λ and are the variations of Lamé coefficients along the direction of aniso-
tropy.
7.3. Finite element modeling
By adopting the same methodology as the one presented in Section 4.3, we provide a
closed form expression of the elastic energy of a linear tetrahedron finite element,
1
t T 1

WNL (T ) = Uj B j k Uk + Uj .CjTkl (Uk .Ul )
2 2
j,k j,k,l
1
T
+ Dj klm (Uj .Uk )(Ul .Um ), (7.4)
2
j,k,l,m
where the terms BjTk , CjTkl and DjTklm , called “stiffness parameters”, are given by
• BjTk is a (3 × 3) symmetric matrix (which corresponds to the linear component of
the energy),

36V (T )BjTk = λ(mj ⊗ mk ) + µ (mk ⊗ mj ) + (mj .mk )I3

λ
+ − + µ (a0 ⊗ a0 )(mj ⊗ mk )(a0 ⊗ a0 ),
2
• CjTkl is a vector,
2 λ µ
216 V (T ) CjTkl = mj (mk .ml ) + ml (mj .mk ) + mk (mj .ml )
2 2
λ
+ − + µ (a0 ⊗ a0 )(mj ⊗ mk )(a0 ⊗ a0 )ml ,
2
• and DjTklm is a scalar,
3 λ µ
1296 V (T ) DjTklm = (mj .mk )(ml .mm ) + (mj .mm )(mk .ml )
8 4

1 λ
+ − + µ (a0 .mj )(a0 .mk )(a0 .ml )(a0 .mm ).
4 2
• The last term of each stiffness parameter models the anisotropic behavior of the
material.
The elastic force applied at each vertex pi of tetrahedron T is obtained as the derivation
of the elastic energy WNL (T ) with respect to the displacement pi ,

1
Fi (T ) = BijT Uj + (Uk ⊗ Uj )CjTki + (Uj .Uk )CijT k
2
j j,k

F1i (T ) F2i (T )

+2 DjTkli Ul Utk Uj . (7.5)

j,k,l

F3i (T )
The first term of the elastic force (F1i (T )) corresponds to the linear elastic case presented
in Section 4.4.
7.4. Non-linear tensor–mass model
In this section, we generalize the tensor–mass model introduced in Section 6.1 to the
case of large displacement elasticity. The only changes in the tensor–mass algorithm are
related to the computation of the elastic force Fi applied at vertex i.
In the case of linear elasticity, this force was computed by a first scan of all edges to
compute the terms [Kij ]uj followed by a scan of all vertices to add the terms [Kii ]ui .
TABLE 7.1
Storage of the stiffness parameters on the mesh
Stiffness parameters Tensors Vectors Scalars

distribution
Vertex p Bpp C ppp D pppp

Edge (p, j ) Bpj C ppj C jpp D jppp D jjjp D jpjp
C jjp C pjj D pjjp D jjpp
Triangle (p, j, k) C j kp D j kpp D jpkp D pj kp
C kjp D jj kp D j kjp D kjjp
C pj k D kkjp D kj kp D j kkp
Tetrahedron (p, j, k, l) D j klp D j lkp D kj lp
D kljp D lj kp D lkjp
p
We proposed to apply the same principle to the quadratic term (F2 (T ) of Eq. (7.5))
p
and the cubic term (F3 (T )). The former requires stiffness vectors for vertices, edges
and triangles, and the latter requires stiffness scalars for vertices, edges, triangles and
tetrahedra.
The task of assembling global stiffness parameters is slightly more time consuming
than in the linear case, since 31 parameters must be assembled instead of 2; these para-
meters are presented in Table 7.1.
For vertex, edge and triangle parameters, one needs to add the contributions of all
neighboring tetrahedra. For instance, the vertex rigidity vector C ppp is computed at
vertex p as

T
C ppp = Cppp .
T ∈S (p)
For the 6 scalar parameters Dj klp stored at each tetrahedron, no assembly is required
since there is no other contribution originating from another tetrahedron.
The computation of the elastic force is performed by successively scanning tetrahe-
dra, triangles, edges and vertices of the mesh. When scanning triangles for instance, the
contributions from the three triangles are computed and added to the elastic force of
each of its three vertices. The contribution for each element is summarized in Eq. (7.5).
edge triangle
Fi = Fvertex
i + Fi + Fi + Ftetrahedron
i (7.6)
with
Vertex contribution
pp
B Up
Fvertex = ,
i
+ (Up ⊗ Up ) + 12 (Up .Up )I3 C ppp
+2Dpppp Up Utp Up
Edge contribution
pj
B Uj

edge

+ (Uj ⊗ Up ) + (Uj .Up )I3 C ppj + (Up ⊗ Uj )C jpp
Fi = ,
+ (Uj ⊗ Uj )C jjp + 12 (Uj .Uj )C pjj
edges(p,j )
+ 2 Djppp 2Up Utp Uj + Uj Utp Up + Djjpp Up Utj Uj
jpjp
+ D + Dpjjp Uj Utj Up + Djjjp Uj Utj Uj
Triangle contribution

(Uk ⊗ Uj )C j kp + (Uj ⊗ Uk )C kjp + (Uj .Uk )C pj k
pj kp
triangle

+2 D + Djpkp Uj Utk Up + Uk Utj Up
Fi = ,
+ 2Dj kpp Up Utj Uk
faces(p,j,k)
+ Dkjjp + Dj kjp Uj Utj Uk + Djj kp Uk Utj Uj

+ Dj kkp + Dkj kp Uk Utk Uj + Dkkjp Uj Utk Uk
Tetrahedron contribution
j klp

2 D + Dkj lp Ul Utj Uk
tetrahedron
Fi = .
+ Dj lkp + Dlj kp Uk Utj Ul
tetra(p,j,k,l)
+ Dkljp + Dlkjp Uj Utk Ul
In terms of data structure, the non-linear tensor–mass model requires the addition of
triangles in the mesh topological description. In our case, we chose to store triangles in
a hash table which is hashed by the three indices of its vertices in lexicographic order.
Furthermore, each tetrahedron owns pointers towards its four triangles and reversely,
each triangle owns pointers towards its two neighboring tetrahedra.
During the simulation of resection, tetrahedra are iteratively removed near the ex-
tremities of virtual cavitron instruments. When removing a single tetrahedron, 280 float-
ing point numbers are updated to suppress the tetrahedron contributions to the stiffness
parameters of the surrounding vertices, edges and triangles:
4 ∗ (1 tensor + 1 vector + 1 scalar)

+ 6 ∗ (1 tensor + 4 vectors + 5 scalars)
+ 4 ∗ (3 vectors + 9 scalars)
= 280 real numbers.
By locally updating stiffness parameters, the tissue has exactly the same properties as if
the corresponding tetrahedron had been removed at its rest position. Because of the vol-
umetric continuity of finite element modeling, the tissue deformation remains realistic
during cutting.
F IG . 7.3. Penalization of volume variation.
7.5. Incompressibility constraint
Living tissue, which is made essentially of water is almost incompressible, a property

which is difficult to model and which, in most cases, leads to instability problems. This
is the case with the St Venant–Kirchhoff model: the material remains incompressible
when the Lamé constant λ tends towards infinity. Taking a large value for λ would
impose to decrease the time step and therefore to increase the computation time. An-
other reason to add an external incompressibility constraint to the model is intrinsic to
the model itself: the St Venant–Kirchhoff model relies on the Green–Lagrange strain
tensor E which is invariant with respect to rotations. But it is also invariant with re-
spect to symmetries, which could lead to the reversal of some tetrahedra under strong
constraints.
We chose to penalize volume variation by applying to each vertex of the tetrahedron
a force directed along the normal of the opposite face Np (see Fig. 7.3), the norm of the
force being proportional to the square of the relative volume variation,

p V − V0 2
Fincomp = sign(V − V0 ) Np . (7.7)
V0
Since the volume V is proportional to the height of each vertex facing its opposite
p
triangle, when V is greater than V0 then the force Fincomp tends to decrease V by moving
each vertex along the normal of the triangle facing it. These forces act as an artificial
pressure inside each tetrahedron. This method is closely related to Lagrange multipliers,
which are often used to solve problem of energy minimization under constraints.
7.6. Results
In a first experiment, we wish to highlight the contributions of our new deformable

model in the case of partial rotations. Fig. 7.4 shows the same experience as the one
presented for linear elasticity (Section 7.1, Fig. 7.2). On the left we can see that the
cylinder vertices are now able to follow non-straight trajectories (Fig. 7.4(a)), leading
F IG . 7.4. (a) Successive deformations of the non-linear model (P ICINBONO , D ELINGETTE and AYACHE
[2003]). Side (b) and top (c) view of the comparison between linear (wireframe) and non-linear model (solid
rendering).
to much more realistic deformations than in the linear (wireframe) case (Figs. 7.4(b)
and (c)).
The second example presents the differences between isotropic and anisotropic ma-
terials. The three cylinders of Fig. 7.5 have their top and bottom faces fixed, and are
submitted to the same forces. While the isotropic model on the left undergoes a “snake-
like” deformation, the last two, which are anisotropic along their height, stiffen in order
to minimize their stretch in the anisotropic direction. The rightmost model, being twice
as stiff as the middle one in the anisotropic direction, starts to squeeze in the plane of
isotropy because it cannot stretch anymore.
In the third example (Fig. 7.6), we apply a force to the right lobe of the liver (the
liver is fixed in a region near the center of its back side, and Lamé coefficients are:
F IG . 7.5. Shearing deformation of tubular structures under the action of the force indicated by the arrow.
The leftmost figure corresponds to an isotropic non-linear material while the center and rightmost figures
correspond to a non-linear anisotropic material, the direction of anisotropy being the cylinder axis.
F IG . 7.6. Linear (upper mesh in wireframe), non-linear (Gauraud shaded) liver models and rest shape (lower
mesh in wireframe). In both cases, the same forces showed in solid lines are applied to three surface nodes
lying on the left lobe (P ICINBONO , D ELINGETTE and AYACHE [2003]).
λ = 40 kPa and µ = 10 kPa). Using the linear elastic model, the right part of the liver
undergoes a large (and unrealistic) volume increase, whereas with non-linear elasticity,
the right lobe is able to rotate partially, while adopting a more realistic deformation.
Adding the incompressibility constraint on the same examples decreases the volume
variation even more (see Table 7.2), and also stabilizes the behaviour of the deformable
models in highly constrained areas.
The last example is the simulation of a typical laparoscopic surgical gesture on the
liver. One tool is pulling the edge of the liver sideways while a bipolar cautery device
TABLE 7.2
Volume variation results. For the cylinder: left, middle and right stand for the different
deformations of Figs. 7.5 and 7.6
Volume variation (%) Linear Non-linear Non-linear incomp.
Cylinder (left – middle – right) 7 – 28 – 63 0.3 – 1 – 2 0.2 – 0.5 – 1

Liver 9 1.5 0.7
F IG . 7.7. Simulation of laparoscopic liver surgery.
cuts it. During the cutting, the surgeon pulls away the part of the liver he wants to
remove. This piece of liver undergoes large displacements and the deformation appears
fairly realistic with this new non-linear deformable model (Fig. 7.7).
Obviously, the computation time of this model is larger than for the linear model
because the force equation is more complex (Eq. (7.5) in Section 7.3 to be compared
with Eq. (6.3) in Section 6.1.7). With our current implementation, the simulation refresh
rate is five times slower than with the linear model. Nevertheless, with this non-linear
model, we can reach an update cycle of 25 Hz on meshes made of about 2000 tetrahedra
(on a PC Pentium PIII 500 MHz). This is enough to achieve real-time visual feedback
with quite complex objects, and even to provide a realistic haptic feedback using force
extrapolation as described in P ICINBONO , L OMBARDO , D ELINGETTE and AYACHE
[2000].
7.7. Optimization of non-linear deformations
We showed that non-linear elasticity allows to simulate much more realistic deforma-
tions than linear elasticity when the model undergoes large displacements. However,
F IG . 7.8. Adaptable non-linear model deformation compared to its rest position (wireframe).
F IG . 7.9. Deformation of the adaptive non-linear model for several values of the threshold.
F IG . 7.10. Updating frequencies of the adaptable model for several values of the threshold.
non-linear elasticity is more computationally expensive than linear elasticity. Since non-
linear elastic forces tend to linear elastic forces as the maximum vertex displacement
decreases to zero, we propose to use non-linear elasticity only at parts of the mesh
where displacements are larger than a given threshold, the remaining part using linear
elasticity. Thus, we modified the force computation algorithm in the following manner:
for each vertex, we first compute the linear part of the force, and we add the non-linear
part only if its displacement is larger than a threshold. Fig. 7.8 shows a deformation
computed with this optimization (same model as in Fig. 7.6). This liver model is made
of 6342 tetrahedra and 1394 vertices. The threshold is set to 2 cm while the mesh is
about 30 cm long. The points drawn on the surface identify vertices using non-linear
elasticity. With this method, we reach an update frequency of 20 Hz instead of 8 Hz
for a fully non-linear model. The same deformation is presented on Fig. 7.9 for differ-
ent values of the threshold. With this method, we can choose a trade-off between the
bio-mechanical realism of the deformation and the update frequency of the simulation.
The diagram on Fig. 7.10 shows the update frequencies reached for each value of the
threshold, in comparison with the fully linear and the fully non-linear models. Even
when this threshold tends towards infinity, the adaptable model is slower than the linear
model, because the computation algorithm of the non-linear force is more complex. In-
deed, the computation of non-linear forces requires to visit all vertices, edges, triangles
and tetrahedra of the mesh, whereas only vertices and edges need to be visited for the
F IG . 7.11. Simulation of hepatectomy based on a non-linear adaptable elastic model. Non-linear elastic force
are applied on vertices outlined with a box.
linear model. For the simulation example of Fig. 7.7, this optimization leads to update
frequencies varying between 50 and 80 Hz, depending on the number of points model-
ing non-linear elasticity (Fig. 7.11). The minimal frequency of 50 Hz is reached at the
end of the simulation, when all vertices of the resected part of the liver are using large
displacement elasticity (on the right of Fig. 7.11).
In general, two strategies can be used to set the value of this threshold. In the first
strategy, the threshold is increased until a given update frequency is matched as demon-
strated previously. The second strategy is physically-motivated and sets the threshold to
10% of the typical size of the mesh since it corresponds to the extent of displacement
for which linear elasticity remains a valid constitutive law.
8. Conclusion
In this chapter, we have presented several algorithms for computing in real-time the de-
formation of soft tissues in a surgical simulator. We wish to stress two important aspects
of these algorithms. First of all, using linear tetrahedra as finite elements helped us to
write closed-form expressions of the elastic energy and its derivatives, even in the case
of large displacement elasticity. These expressions nicely decouple the physical parame-
ters (Lamé coefficients) from the geometry of each tetrahedron both in its rest position
(direction of anisotropy, rest volume, area vectors) and in its deformed state (displace-
ment vectors). Furthermore, it enables to quickly assemble local and global stiffness
matrices when the mesh topology has been modified during a cutting simulation.
Second, in the context of surgery simulation, soft tissue deformation algorithms are
closely tied with the visualization, collision detection and haptic rendering algorithms.
Furthermore, the traditional stages of matrix assembly, matrix preconditioning, system
solution and post-processing, cannot be easily decoupled like in classical software pack-
ages available in structural mechanics. This implies that the data structure and the flow
chart must be carefully designed in order to achieve a reasonable trade-off between
these performances. Therefore, building a successful simulator can only be achieved
by a multidisciplinary effort covering the fields of biomechanics, numerical analysis,
robotics and computer graphics.
An hepatectomy simulator based on the quasi-static precomputed linear elastic model
(introduced in Section 5) and the large displacement non-linear elastic models (intro-
duced in Section 7) has been built where the following three basic surgical gestures can
be rehearsed: touching soft tissue, gripping soft tissue and cutting parenchyma with a
cavitron. Furthermore, we recently added a physical model of the portal vein (F OREST,
D ELINGETTE and AYACHE [2003]), which allows the user to simulate the clamping and
cutting of vessels during the hepatic resection.
However, to increase the training impact and realism of the simulation, it is important
to simulate the contact between the liver and neighboring structures such as the gall-
bladder, the different ligaments, the right kidney, the peritoneum, etc. These additional
surface and volumetric models require to extend the soft tissue models introduced in
this chapter in two ways.
First, it is necessary to extend the precomputed linear elastic model to include large-
displacement non-linear elasticity. Indeed, the linear domain of biological soft tissue is
usually rather small, and therefore many surgical gestures can only be simulated by us-
ing large-displacement elasticity (like rotating the lobe of the liver or resecting the gall-
bladder). The precomputation of non-linear elastic material is not a trivial task since
it implies solving a complex third-order algebraic equation in the case of St Venant–
Kirchhoff elasticity (see Section 7.2). Instead, it may be possible to find suitable ap-
proximations which can be computed efficiently.
Second, it is necessary to extend the concept of hybrid models (introduced in Sec-
tion 6.3) in order to cope with the deformation of models including several tens of
thousands of vertices. Ideally, we would like to provide accurate but computationally
expensive soft tissue models in the center of the surgical field where the user performs
complex gestures and at the same time to provide less expensive models but potentially
less accurate, away from the center of the surgical field. Of course, during surgery, the
focus of the surgeon may switch from the gall-bladder to the hepatic parenchyma which
implies that those tissue models should evolve dynamically from one level of accuracy
to the other. Achieving this level of scalability with the constraint that the topology of
these models may change over time, is the main challenge of soft tissue modeling for
surgery simulation.
Finally, we would like to stress the importance of validating the different components
of a surgical simulator. Concerning soft tissue models, there are at least three levels of
validation that need to be achieved. A first validation consists in comparing the soft
tissue deformation algorithms that rely on strong hypotheses against well-known finite
element packages in order to evaluate the range of approximations that are performed.
In the second level of validation, the biomechanical behavior of each anatomical struc-
ture must be compared to experimental dataset. Ideally, one would like to validate both
boundary conditions and the constitutive law of each biological tissue. However, in
practice, this validation is made difficult by the lack of quantitative experimental in-
formation. The third level of validation consists in evaluating the dynamic behaviour of
each soft tissue during the simulation since some models that appear too soft or too stiff.
Finally, and most importantly, it is required to validate the whole simulation system by
assessing its ability to succeed in training young residents to perform a given surgical
task.
Despite these remaining issues to be solved, we believe that practical surgery simula-
tors will be fully operational and actually part of the surgical studies in the near future.
List of mathematical symbols
fu Update frequency of the soft tissue model

t Discrete or continuous time variable
Xt Position of the model at time t
Trelaxation Relaxation time
Tc Computation time
Tinteraction Latency caused by the software and hardware architecture
t Time step used in the discretization of temporal derivatives
F Global force vector
K Global stiffness matrix
U Global displacement vector

M Global mass matrix
C Global damping matrix
U̇ Global speed vector
Mrest Soft tissue model at its rest position
Mdef Soft tissue model at its deformed position
Ω Region of space for the rest configuration
Φ(x, y, z) Deformation function that maps point (x, y, z) from the rest configuration
to the deformed configuration
X Point in the rest configuration
U(X) Displacement function
C(X) Right Cauchy–Green strain tensor
E(X) Green–Lagrange strain tensor
I3 3 × 3 identity matrix
EL Linearized strain tensor
eij Element of the linearized strain tensor
T(X) Cauchy stress tensor
W (X) Density of elastic energy
λ, µ Isotropic Lamé coefficients
E, ν Isotropic Young modulus and Poisson ratio
a0 Unit vector along the direction of anisotropy for transversally isotropic materials
λa0 , µa0 Lamé coefficients along the direction of anisotropy
λ, µ Difference between the Lamé coefficients along the direction of anisotropy
and those in the orthogonal plane
WAni (X) Additional term of the density of elastic energy caused by anisotropy
I4 , I5 Deformation invariants estimated along the direction of anisotropy
pi Point of a tetrahedron in its rest position
qi Point of a tetrahedron in its deformed position
ui Displacement vector of a vertex of a tetrahedron
T Tetrahedron as a linear finite element
hj (X) Shape functions associated with a linear tetrahedron
P 4 × 4 matrix describing the shape functions
V (T ) Volume of tetrahedron T
mi Area vector opposite to vertex i
Vi 6 times the volume of the tetrahedron made by the origin o and vertices pi+1 , pi+2
and pi+3
Ti Triangle opposite to vertex i
ni Normal vector at the triangle Ti opposite to vertex i in a tetrahedron
θi,j Angle between normal vectors of triangles Ti and Tj
Ai Area of triangle Ti
li,j Length of the edge connecting vertices i and j
fi Height of vertex o above triangle Ti
BTi,j Element (i, j ) of the 3 × 3 stiffness matrix for a tetrahedron T made of an isotropic
material
ATi,j Element (i, j ) of the 3 ×3 stiffness matrix for a tetrahedron T made of an transver-
sally isotropic material
Ki,j 3 × 3 global stiffness matrix between vertex i and j
ki,j Eigenvalue along the edge direction of matrix Ki,j
Wg (T ) Work of gravity forces
Wp (T ) Work of external surface pressure
Mi,j 3 × 3 global mass matrix between vertex i and j
K i,j 3 × 3 global stiffness matrix between vertex i and j that includes spring boundary
conditions
Rg Global vector of gravity forces
Rb Global vector of boundary forces
Acknowledgements
We thank Matthias Teschner, Denis Laurendeau and Jean-Marc Schwartz for their price-
less comments and for proofreading this article.
The work presented in this paper is a joint work between the authors and mainly
two former PhD students: Stéphane Cotin and Guillaume Picinbono. Stéphane Cotin
developed the precomputed linear elastic model of Section 5 as well as a first version
of the tensor–mass model described in Section 6.1. Guillaume Picinbonno proposed
the extension of the tensor–mass model to the case of large displacement elasticity (in
Section 7). We also thank Clément Forest and Jean-Christophe Lombardo for their nu-
merous contributions on force-feedback rendering, collision detection as well as mesh
data structure. This work was fueled with the stimulating remarks and propositions from
our INRIA colleagues who participated in the AISIM and CAESARE joint initiatives:
Marie-Paule Cani, Marina Vidrascu, Marc Thiriet, Christian Laugier. Also, we are grate-
ful to Prof. Marescaux, Prof. Leroy and Prof. Luc Soler from the IRCAD research center
for their long-term vision and for sharing their expertise of abdominal surgery with us.
Finally, we acknowledge the strong support we received from Gilles Khan, INRIA Vice-
President for Research, during the different stages of this research work.
References
A CKERMAN , M.J. (1998). The visible human project. Proc. IEEE: Special Issue on Surgery Simula-
tion 86 (3), 504–511.
AYACHE , N. (2003). Epidaure: a research project in medical image analysis, simulation and robotics at IN-
RIA. IEEE Trans. Medical Imaging, Invited Editorial.
AYACHE , N., D ELINGETTE , H. (eds.) (2003). Int. Symp. on Surgery Simulation and Soft Tissue Modeling,
Juan-Les-Pins, France, June 1998. In: Lecture Notes in Comput. Sci. 2673 (Springer-Verlag, New York).
BARAFF , D., W ITKIN , A. (1998). Large steps in cloth simulation. In: Computer Graphics Proceedings,
SIGGRAPH’98, Orlando, USA, July 1998, pp. 43–54.
BASSINGTHWAIGHTE , J.B. (2000). Strategies for the physiome project. Ann. Biomed. Engrg. 28, 1043–1058.
BATHE , K.-L. (1982). Finite Element Procedures in Engineering Analysis (Prentice Hall, New York).
BAYRAKTAR , H., A DAMS , M., G UPTA , A., PAPADOPOULOS , P., K EAVENY, T. (2003). The role of large de-
formations in trabecular bone mechanical behavior. In: ASME Bioengineering Conference, Key Biscayne,
FL, USA, June 2003.
B ENZLEY , S.E., P ERRY , E., C LARK , B., M ERKLEY , K., S JAARDEMA , G. (1995). Comparison of all-
hexahedral and all-tetrahedral finite element meshes for elastic and elasto-plastic analysis. In: Proc. 4th
Int. Meshing Roundtable, Sandia National Laboratories, October 1995, pp. 179–191.
B ERCI , G., H UNTER , J.G., S ACKIER , J.M. (1994). Training in laparoscopic cholecystectomy: Quantifying
the learning curve. J. Endoscopic Surgery 8, 28–31.
B ESAG , J. (1986). On the statistical analysis of dirty pictures. J. Roy. Statist. Soc. 48 (3), 326–338.
B IELSER , D., G ROSS , M.H. (2000). Interactive simulation of surgical cuts. In: Proc. Pacific Graphics 2000,
Hong-Kong, October 2000 (IEEE Computer Society Press), pp. 116–125.
B OUX DE C ASSON , F., L AUGIER , C. (1999). Modelling the dynamics of a human liver for a minimally
invasive simulator. In: Proc. Int. Conf. on Medical Image Computer-Assisted Intervention, Cambridge,
UK, September 1999.
B RO -N IELSEN , M. (1998). Finite element modeling in surgery simulation. Proc. IEEE: Special Issue on
Surgery Simulation 86 (3), 490–503.
B RO -N IELSEN , M., C OTIN , S. (1996). Real-time volumetric deformable models for surgery simulation using
finite elements and condensation. In: Eurographics’96, vol. 3, pp. 57–66.
B RONSHTEIN , I.N., S EMENDYAYEV , K.A. (1985). Handbook of Mathematics (Van Nostrand–Reinhold,
New York).
B ROWN , J.D., ROSEN , J., K IM , Y., C HANG , L., S INANAN , M., H ANNAFORD , B. (2003). In-vivo and in-situ
compressive properties of porcine abdominal soft tissue. In: Medicine Meets Virtual Reality, MMVR’03,
Newport Beach, USA, January 2003.
C ANAS , J., PARIS , F. (1997). Boundary Element Method: Fundamentals and Application (Oxford Univ.
Press, London).
C ARTER , F.J. (1998). Biomechanical testing of intra-abdominal soft tissue. In: Int. Workshop on Soft Tissue
Deformation and Tissue Palpation, Cambridge, MA, October 1998.
C IARLET , P.G. (1987). Mathematical Elasticity, vol. 1: Three-dimensional Elasticity (North-Holland, Ams-
terdam). ISBN 0-444-70259-8.
C OSMI , F. (2001). Numerical solution of plane elasticity problems with the cell method. Comput. Methods
Engrg. Sci. 2 (3).
C OSTA , I.F., BALANIUK , R. (2001). Lem – an approach for real time physically based soft tissue simulation.
In: Int. Conf. Automation and Robotics, ICRA’2001, Seoul, May 2001.
546
References 547
C OTIN , S., D ELINGETTE , H., AYACHE , N. (1999a). Real-time elastic deformations of soft tissues for surgery
simulation. IEEE Trans. Visual. Comput. Graph. 5 (1), 62–73.
C OTIN , S., D ELINGETTE , H., AYACHE , N. (1999b). Real-time elastic deformations of soft tissues for surgery
simulation. IEEE Trans. Visual. Comput. Graph. 5 (1), 62–73.
C OTIN , S., D ELINGETTE , H., AYACHE , N. (2000). A hybrid elastic model allowing real-time cutting, defor-
mations and force-feedback for surgery training and simulation. The Visual Computer 16 (8), 437–452.
C OTIN , S., D ELINGETTE , H., C LEMENT , J.-M., TASSETTI , V., M ARESCAUX , J., AYACHE , N. (1996).
Volumetric deformable models for simulation of laparoscopic surgery. In: Proc. Int. Symp. on Computer
and Communication Systems for Image Guided Diagnosis and Therapy, Computer Assisted Radiology,
CAR’96. In: Int. Congr. Ser. 1124 (Elsevier, Amsterdam).
C OUINAUD (1957). Le foie, Études anatomiques et chirurgicales (Masson, Paris).
C OVER , S.A., E ZQUERRA , N.F., O’B RIEN , J.F. (1993). Interactively deformable models for surgery simu-
lation. IEEE Comput. Graph. Appl. 13, 68–75.
D AN D. (1999). Caractérisation mécanique du foie humain en situation de choc, PhD thesis, Université
Paris 7.
D EBUNNE , G., D ESBRUN , M., C ANI , M.-P., BARR , A.H. (2001). Dynamic real-time deformations using
space and time adaptive sampling. In: Computer Graphics Proceedings, SIGGRAPH’01, August 2001.
D ECORET , X., S CHAUFLER , G., S ILLION , F., D ORSEY , J. (1999). Multi-layered impostors for accelerated
rendering. Computer Graphics Forum (Eurographics’99) 18, 61–73.
D ELINGETTE , H. (1998). Towards realistic soft tissue modeling in medical simulation. Proc. IEEE: Special
Issue on Surgery Simulation 86, 512–523.
D ELINGETTE , H. (1999). General object reconstruction based on simplex meshes. Int. J. Comput. Vi-
sion 32 (2), 111–146.
D ELINGETTE , H., M ONTAGNAT , J. (2001). Shape and topology constraints on parametric active contours.
J. Comput. Vision and Image Understanding 83, 140–171.
D ESBRUN , M., G ASCUEL , M.-P. (1995). Animating soft substances with implicit surfaces. In: Computer
Graphics, SIGGRAPH’95, Los Angeles.
D EUSSEN , O., K OBBELT , L., T UCKE , P. (1995). Using simulated annealing to obtain a good approxima-
tion of deformable bodies. In: Proc. Eurographics Workshop on Animation and Simulation, Maastricht,
Netherlands, September 1995 (Springer-Verlag, Berlin).
D UNCAN , J., AYACHE , N. (2000). Medical image analysis: Progress over two decades and the challenges
ahead. IEEE Trans. on Pattern Analysis and Machine Intelligence 22 (1), 85–106.
F OREST , C., D ELINGETTE , H., AYACHE , N. (2002a). Cutting simulation of manifold volumetric meshes.
In: Modelling and Simulation for Computer-aided Medicine and Surgery, MS4CMS’02.
F OREST , C., D ELINGETTE , H., AYACHE , N. (2002b). Removing tetrahedra from a manifold mesh. In: Com-
puter Animation, CA’02, Geneva, Switzerland, June 2002 (IEEE Computer Society), pp. 225–229.
F OREST C., D ELINGETTE H., AYACHE N. (2003). Simulation of surgical cutting in a manifold mesh by
removing tetrahedra, Medical Image Analysis, submitted for publication.
F RANCE , L., A NGELIDIS , A., M ESEURE , P., C ANI , M.-P., L ENOIR , J., FAURE , F., C HAILLOU , C. (2002).
Implicit representations of the human intestines for surgery simulation. In: Conf. on Modeling and Simu-
lation for Computer-aided Medicine and Surgery, MS4CMS’02, Rocquencourt, November 2002.
F RANCE , L., L ENOIR , J., M ESEURE , P., C HAILLOU , C. (2002). Simulation of minimally invasive surgery
of intestines. In: Richir, S. (ed.), Fourth Virtual Reality International Conference, VRIC’2002, pp. 21–27.
ISBN 2-9515730.
F UNG , Y.C. (1993). Biomechanics – Mechanical Properties of Living Tissues, second ed. (Springer-Verlag,
Berlin).
G IBSON , S., S AMOSKY , J., M OR , A., F YOCK , C., G RIMSON , E., K ANADE , T., K IKINIS , R., L AUER , H.,
M C K ENZIE , N. (1997). Simulating arthroscopic knee surgery using volumetric object representations,
real-time volume rendering and haptic feedback. In: Troccaz, J., Grimson, E., Mosges, R. (eds.), Proc.
First Joint Conf. CVRMed-MRCAS’97. In: Lecture Notes in Comput. Sci. 1205, pp. 369–378.
G LADILIN E. (2002). Biomechanical modeling of soft tissue and facial expressions for craniofacial surgery
planning, PhD thesis, Freie Univerisität Berlin, Germany.
G OTTSCHALK , S., L IN , M., M ANOCHA , D. (1996). Obb-tree: A hierarchical structure for rapid interference
detection. In: Proc. SIGGRAPH 96, New Orleans, LA, pp. 171–180. ISBN 0-201-94800-1.
H ODGSKINSON , R., C URREY , J.D. (1992). Young modulus, density and material properties in cancellous
bone over a large density range. J. Materials Science: Materials in Medicine 3, 377–381.
H OUBOLT , J.C. (1950). A recurrence matrix solution for the dynamic response of elastic aircraft. J. Aero-
nautical Sci. 17, 540–550.
H UMPHREY , J.D., S TRUMPF , R.K., Y IN , F.C.P. (1990). Determination of a constitutive relation for passive
myocardium: I. A new functional form. ASME J. Biomech. Engrg. 112, 333–339.
H UMPHREY , J.D., Y IN , F.C.P. (1987). On constitutive relations and finite deformations of passive cardiac
tissue: I. A pseudostrain-energy function. ASME J. Biomech. Engrg. 109, 298–304.
H UNTER , P., P ULLAN , A. (1997). FEM/BEM Notes (University of Auckland, New-Zeland). Available at
http://www1.esc.auckland.ac.nz/Academic/Texts/fembemnotes.pdf.
JAMES , D.L., PAI , D.K. (1999). Artdefo accurate real time deformable objects. In: Computer Graphics,
SIGGRAPH’99, pp. 65–72.
J OE , B. (1991). Geompack – a software package for the generation of meshes using geometric algorithms.
J. Advanced Eng. Software 13, 325–331.
K AISS , M., L E TALLEC , P. (1996). La modélisation numérique du contact œil-trépan. Revue Européenne des
éléments Finis 5 (3), 375–408.
K ASS , M., W ITKIN , A., T ERZOPOULOS , D. (1988). Snakes: Active contour models. Int. J. Comput. Vision 1,
321–331.
K AUER , M., V USKOVIC , V., D UAL , J., S ZÉKELY , G., BAJKA , M. (2001). Inverse finite element charac-
terization of soft tissues. In: Proc. 4th Int. Conf. on Medical Image Computing and Computer-Assisted
Intervention, MICCAI’01, Utrecht, October 2001. In: Lecture Notes in Comput. Sci. 2208, pp. 128–136.
K AYE , J., P RIMIANO , F., M ETAXAS , D. (1997). A 3d virtual environment for modeling mechanical car-
diopulmonary interactions. Medical Image Analysis (Media) 2 (2), 1–26.
K NOTT , D., PAI , D. (2003). Collision and interference detection in real-time using graphics hardware. In:
Proc. Graphics Interface, Halifax, Canada, June 2003.
K UHN , C H ., K ÜHNAPFEL , U., K RUMM , H.-G., N EISIUS , B. (1996). A ‘virtual reality’ based training sys-
tem for minimally invasive surgery. In: Proc. Computer Assisted Radiology, CAR’96, Paris, June 1996,
pp. 764–769.
K UHNAPFEL , U., A KMAK , H., M AA , H. (2000). Endoscopic surgery training using virtual reality and de-
formable tissue simulation. Computers and Graphics 24, 671–682.
L IU , Z., B ILSTON , L.E. (2000). On the viscoelastic character of liver tissue: experiments and modelling of
the linear behaviour. Biorheology 37, 191–201.
L IU , Z., B ILSTON , L.E. (2002). Large deformation shear properties of liver tissue. Biorheology 39, 735–742.
L OMBARDO , J.-C., C ANI , M.-P., N EYRET , F. (1999). Real-time collision detection for virtual surgery. In:
Computer Animation, Geneva, Switzerland, May 1999, pp. 82–89.
L ORENSEN , W., C LINE , H.E. (1987). Marching cubes: a high resolution 3d surface construction algorithm.
ACM Computer Graphics (SIGGRAPH’87) 21, 163–169.
L OUCHET , J., P ROVOT , X., C ROCHEMORE , D. (1995). Evolutionary identification of cloth animation model.
In: Workshop on Computer Animation and Simulation, Eurographics’95, pp. 44–54.
L UMSDAINE , A., S IEK , J. (1998). The Matrix Template Library. http://www.lsc.nd.edu/research/mtl/.
M ACMILLAN , R.H. (1955). A new method for the numerical evaluation of determinants. J. Roy. Aeronaut.
Soc. 59 (772).
M ANDUCA , A., M UTHUPILLAI , R., ROSSMAN , P., G REENLEAF , J., E HMAN , L. (1996). Visualization of
tissue elasticity by magnetic resonance elastography. In: Proc. of Visualization in Biomedical Imaging,
VBC’96, Hamburg, Germany, pp. 63–68.
M ARCUS , B. (1996). Hands on: Haptic feedback in surgical simulation. In: Proc. of Medicine Meets Virtual
Reality IV, MMVR IV, San Diego, CA, January 1996, pp. 134–139.
M ARK , W., R ANDOLPH , S., F INCH , M., VAN V ERTH , J., TAYLOR II, R.M. (1996). Adding force feedback
to graphics systems: Issues and solutions. In: Rushmeier, H. (ed.), ACM SIGGRAPH Computer Graphics
Annual Conference, SIGGRAPH’96 (Addison–Wesley, Reading, MA), pp. 447–452.
M AUREL , W., W U , Y., M AGNENAT T HALMANN , N., T HALMANN , D. (1998). Biomechanical Models for
Soft Tissue Simulation, ESPRIT Basic Research Series (Springer-Verlag, Berlin).
S ERRANO , C.M., L AUGIER , C. (2001). Realistic haptic rendering for highly deformable virtual objects. In:
Proc. Int. Conf. on Virtual Reality, Yokohama, Japan, March 2001.
References 549
M ILLER , K. (2000). Constitutive modelling of abdominal organs. J. Biomech. 33 (3), 367–373.

M ONTAGNAT , J., D ELINGETTE , H. (1998). Globally constrained deformable models for 3d object recon-
struction. Signal Processing, 173–186.
N AVA , A., M AZZA , E., K LEINERMANN , F., AVIS , N., M C C LURE , J. (2003). Determination of the mechan-
ical properties of soft human tissues through aspiration experiments. In: Proc. Conf. on Medical Robotics,
Imaging And Computer Assisted Surgery, MICCAI 2003, Montreal, Canada, November 2003. In: Lecture
Notes in Comput. Sci.
N EWMARK , N.M. (1959). A method of computation for structural dynamics. J. Engrg. Mech. Division 85,
67–94.
O’M AHONY , A., W ILLIAMS , J., K ATZ , J. (1999). Anisotropic elastic properties of cancellous bone from a
human edentulous mandible. In: Proc. ASME Bioengineering’99 Conference.
O WEN S. (2000). A survey of unstructured mesh generation technology. Technical report, Department of
Civil and Environmental Engineering, Carnegie Mellon University.
PAPADEMETRIS , X., S HI , P., D IONE , D.P., S INUSAS , A.J., C ONSTABLE , R.T., D UNCAN , J.S. (1999).
Recovery of soft tissue object deformation from 3d image sequences using biomechanical models. In:
XVIth Int. Conf. on Information Processing In Medical Imaging, IPMI’99, Visegrád, Hungary, June 28–
July 2, 1999, pp. 352–357.
PARTHASARATHY , V.N., G RAICHEN , C.M., H ATHAWAY , A.F. (1993). A comparison of tetrahedron quality
measures. Finite Elements in Analysis and Design 15, 255–261.
P ICINBONO , G., D ELINGETTE , H., AYACHE , N. (2001). Non-linear and anisotropic elastic soft tissue models
for medical simulation. In: IEEE Int. Conf. Robotics and Automation, ICRA’2001, Seoul, Korea, May
2001. Best conference paper award.
P ICINBONO , G., D ELINGETTE , H., AYACHE , N. (2003). Non-linear anisotropic elasticity for real-time
surgery simulation. Graphical Models 65 (5), 305–321.
P ICINBONO , G., L OMBARDO , J.-C., D ELINGETTE , H., AYACHE , N. (2000). Anisotropic elasticity and
forces extrapolation to improve realism of surgery simulation. In: IEEE Int. Conf. Robotics and Automa-
tion, ICRA’2000, San Francisco, USA, April 2000, pp. 596–602.
P ICINBONO , G., L OMBARDO , J.-C., D ELINGETTE , H., AYACHE , N. (2002). Improving realism of a surgery
simulator: linear anisotropic elasticity, complex interactions and force extrapolation. J. Visual. Comput.
Animation 13 (3), 147–167.
P RESS , W.H., F LANNERY , B.P., T EUKOLSKY , S.A., V ETTERLING , W.T. (1991). Numerical Recipes in C
(Cambridge Univ. Press, Cambridge, UK).
P RESS , W.H., F LANNERY , B.P., T EUKOLSKY , S.A., V ETTERLING , W.T. (1992). Numerical Recipes in
FORTRAN: The Art of Scientific Computing, second ed. (Cambridge Univ. Press, Cambridge, UK).
P USO , M.A., W EISS , J.A. (1998). Finite element implementation of anisotropic quasi-linear viscoelasticity
using a discrete spectrum approximation. ASME J. Biomech. Engrg. 120 (1).
P UTTI , M., C ORDES , C. (1998). Finite element approximation of the diffusion operator on tetrahedra. SIAM
J. Scientific Comput. 19 (4), 1154–1168.
Q UARTERONI , A., T UVERI , M., V ENEZIANI , A. (2000). Computational vascular fluid dynamics: problems,
models and methods. Computing and Visualization in Science 2, 163–197.
R ADETZKY , A. (1998). The simulation of elastic tissues in virtual medicine using neuro-fuzzy systems. In:
Medical Imaging’98: Image Display, San Diego, CA, February 1998.
S AAD , Y. (1996). Iterative Methods for Sparse Linear Systems (WPS).
S AKUMA , I., N ISHIMURA , Y., K ONG C HUI , C., K OBAYASHI , E., I NADA , H., C HEN , X., H ISADA , T.
(2003). In vitro measurement of mechanical properties of liver tissue under compression and elongation
using a new test piece holding method with surgical glue. In: Int. Symp. on Surgery Simulation and Soft
Tissue Modeling, Juan-Les-Pins, France, June 2003. In: Lecture Notes in Comput. Sci. 2673 (Springer-
Verlag, Berlin), pp. 284–292.
S ATAVA , R. (1994). Medicine 2001: The King Is Dead. In: Proc. Conf. Virtual Reality in Medicine.
S ATAVA , R. (1996). Medical virtual reality: The current status of the future. In: Proc. 4th Conf. Medicine
Meets Virtual Reality, MMVR IV, pp. 100–106.
S CHROEDER , W.J., Z ARGE , J., L ORENSEN , W. (1992). Decimation of triangles meshes. Computer Graphics
(SIGGRAPH’92) 26.
S ERMESANT , M., C OUDIÈRE , Y., D ELINGETTE , H., AYACHE , N. (2002). Progress towards an electro-
mechanical model of the heart for cardiac image analysis. In: IEEE Int. Symp. Biomedical Imaging,
ISBI’02, pp. 10–14.
S ERMESANT , M., FARIS , O., E VANS , F., M C V EIGH , E., C OUDIÈRE , Y., D ELINGETTE , H., AYACHE , N.
(2003). Preliminary validation using in vivo measures of a macroscopic electrical model of the heart. In:
Ayache, N., Delingette, H. (eds.), Int. Symp. Surgery Simulation and Soft Tissue Modeling, IS4TM’03. In:
Lecture Notes in Comput. Sci. 2673 (Springer-Verlag, Heidelberg).
S ILLION , F.X., D RETTAKIS , G., B ODELET , B. (1997). Efficient impostor manipulation for real-time visual-
ization of urban scenery. In: Proc. Eurographics’97, Budapest, Hungary, September 1997.
S IMAIL: product of Simulog S.A. – 1, rue James Joule, 78286 Guyancourt cedex, France, http://www.
simulog.fr.
S OFERMAN , Z., B LYTHE , D., J OHN , N. (1998). Advanced graphics behind medical virtual reality: Evolution
of algorithms, hardware and software interfaces. Proc. IEEE: Special Issue on Surgery Simulation 86 (3),
531–554.
S OLER , L., D ELINGETTE , H., M ALANDAIN , G., M ONTAGNAT, J., AYACHE , N., C LÉMENT, J.-M., K OEHL ,
C., D OURTHE , O., M UTTER , D., M ARESCAUX , J. (2000). Fully automatic anatomical, pathological and
functional segmentation from ct-scans for hepatic surgery. In: Medical Imaging 2000, San Diego, February
2000.
S OLER , L., M ALANDAIN , G., D ELINGETTE , H. (1998). Segmentation automatique: application aux an-
gioscanners 3d du foie. Traitement du signal 15 (5), 411–431 (in French).
S PENCER , A.J.M. (1972). Deformations of Fibre-Reinforced Materials (Clarendon Press, Oxford).
S PENCER , A.J.M. (1984). Continuum Theory of Fiber-Reinforced Composites (Springer-Verlag, New York).
S ZEKELY , G., BAIJKA , M., B RECHBUHLER , C. (1999). Virtual reality based simulation for endoscopic
gynaecology. In: Proc. Medicine Meets Virtual Reality, MMVR’99, San Francisco, USA, pp. 351–357.
T ESCHNER , M., H EIDELBERGER , B., M ULLER , M., P OMERANETS , D., G ROSS , M. (2003). Optimized
spatial hashing for collision detection of deformable objects. In: Proc. Vision, Modeling, Visualization,
VMV’03, Munich, Germany, November 2003.
V IDRASCU , M., D ELINGETTE , H., AYACHE , N. (2001). Finite element modeling for surgery simulation. In:
First MIT Conf. on Computational Fluid and Solid Mechanics.
V LACHOS , A., P ETERS , J., B OYD , C., M ITCHELL , J.L. (2001). Curved pn triangles. In: 2001 ACM Symp.
on Interactive 3D Graphics.
W EISS , J.A., G ARDINER , J.C., Q UAPP , K.M. (1995). Material models for the study of tissues mechanics.
In: Proc. Int. Conf. on Pelvic and Lower Extremity Injuries, Washington, DC, December 1995, pp. 249–
261.
W OO , M., N EIDER , J., D AVIS , T. (1997). OpenGL Programing Guide (Addison–Wesley, Reading, MA).
YAMASHITA , Y., K UBOTA , M. (1994). Ultrasonic characterization of tissue hardness in the in-vivo human
liver. In: Proc. IEEE Ultrasonics Symposium, pp. 1449–1453.
Z ORIN , D., S CHROEDER , P., S WELDENS , W. (1996). Interpolating subdivision for meshes with arbitrary
topology. In: Proc. 23rd Annual Conf. on Computer Graphics and Interactive Techniques (ACM Press),
pp. 189–192.
Recovering Displacements and
Deformations from 3D Medical Images
Using Biomechanical Models
Xenophon Papademetris
Departments of Biomedical Engineering and Diagnostic Radiology, Yale University,
New Haven, CT, USA
E-mail: xenophon.papademetris@yale.edu
Oskar Škrinjar
Department of Biomedical Engineering, Georgia Institute of Technology,
Atlanta, GA, USA
E-mail: oskar.skrinjar@bme.gatech.edu
James S. Duncan
Departments of Electrical Engineering, Biomedical Engineering and
Diagnostic Radiology, Yale University, New Haven, CT, USA
E-mail: james.duncan@yale.edu
1. Introduction
The primary emphasis of this chapter is to describe the use of biomechanical models for
the estimation of non-rigid displacement fields from sequences of three-dimensional
medical images. In both case studies described later in this chapter, namely (i) the esti-
mation of brain shift for neurosurgery and (ii) the estimation of left ventricular deforma-
tion, the proper modeling of the underlying tissue is important in order to ensure reliable
and robust estimation of the underlying displacement and consequently the deforma-
tion. Modeling is needed as the image-derived displacement estimates generated from
a number of methods (to be described in Section 2) have the following characteristics:
551
552 X. Papademetris et al.
• They are sparse. Displacements are only available at certain points and not the
whole of the material.
• They are noise-corrupted. This is an inherent problem in all medical image analysis
methods, although the level of noise is very method dependent.
• They may contain only partial information. Even where displacements are avail-
able, only a certain component of the displacement vector may be known.
The selection of an appropriate model and an appropriate modeling framework are
of great importance for the estimation of complete and smooth displacement fields. The
rest of this chapter reads as follows.
In Section 2 we present the underlying mathematical framework for the use of con-
tinuum mechanical models within this image analysis context. In particular, we present
a brief introduction to continuum mechanics (Section 2.1) followed by a description
of two frameworks for the integration of image-derived information with a mechani-
cal model. The section continues with a description of a new and unique continuum
mechanical model, the active elastic model devised specifically to model actively de-
forming tissue, which we will later demonstrate in Section 4. In the final part of this
section we describe the key numerical technique used in this chapter – the finite element
method.
We use two case studies to illustrate the use of this underlying mathematical frame-
work. In Section 3 we describe methodology to compensate for brain shift in image
guided neurosurgery and in Section 4 we describe algorithms to estimate the deforma-
tion of the left ventricle of the heart. Both sections are structured as follows. First, the
background of the problems is given, followed by the overall system design (typically
a sequence of image segmentation and mesh generation followed by image-based dis-
placement data extraction). Next, the specifics of the mechanical model are presented,
followed by validation results on real and simulated data.
We conclude the chapter with some further thoughts and remarks in Section 5.
2. Mathematical framework
In this section we introduce key concepts from continuum mechanics (Section 2.1) and
present two possible frameworks for the integration of such material models with image-
derived information (Sections 2.2 and 2.3.) In Section 2.4 we derive a modification
of the linear elastic model, the active elastic model designed specifically to account
for materials which are undergoing active rather than passive deformation.1 Finally,
in Section 2.5 we present an overview of the finite element method – the numerical
technique used to solve for the displacements/deformations in the application of this
methodology.
1 We use the term passive deformation to refer to the change in material shape caused by external forces,
e.g., gravity. The term active deformation is used to describe the change in shape caused by the object itself,
e.g., muscle contraction.
Recovering displacements and deformations from 3D medical images 553
2.1. A brief introduction to continuum mechanics
2.1.1. Deformations
The deformation gradient matrix. In this section we follow the presentations in
S PENCER [1980], Chapter 6, and H UNTER , NASH and S ANDS [1997]. Consider a body
B(0), shown in Fig. 2.1, which after time t moves and deforms to body B(t). A material
particle initially located at some position X on B(0) moves to a new position x on B(t).
If we further assume that material cannot appear or disappear there will be a one-to-one
correspondence between X and x, so we can always write the path of the particle as
x = x(X, t). (2.1)
We can also define the displacement vector for this particle as
u(t) = x(t) − X. (2.2)
This relationship is also invertible, given x and t, we can find X. Let us consider two
neighboring particles located at X and X + dX on B(0). In a new configuration B(t)
using Eq. (2.1), we can write
∂x
dx = dX. (2.3)
∂X
The Jacobian matrix F (t) = ∂x(t)/∂X is called the deformation gradient matrix. We
note that by definition, F (0) = I . Using this, we can rewrite Eq. (2.1) more fully as
dx(t) = F (t).dX, (2.4)



∂xi
F (0) = I,

 ∂Xj ,
1, i = j,
Fij = δij =

 ∂ui 0, otherwise.

 + δij , u(0) = 0,
∂Xj
The mapping defined by Eqs. (2.1)–(2.4) has two components: a rigid motion com-
ponent and a change in the shape or deformation of the object. For the purposes of cap-
turing the material behavior (to be discussed in Section 2.1), we need to extract from
F IG . 2.1. Definition of displacement. Figure reprinted from PAPADEMETRIS , S INUSAS , D IONE and D UN -
CAN [2001], Estimation of 3D left ventricular deformation from echocardiography, Medical Image Analysis
5(1):17–29, ©2001 by permission from Elsevier.
F the component which is a function of the rigid motion and the component which is a
function of the deformation.
To extract the deformation component we use the polar decomposition (S TRANG
[1986]) to write F as
F= R
× U
. (2.5)
Rotation matrix Symmetric matrix
The matrix R is a rotation matrix having the properties R ∗ R t = I , det(R) = 1, and U

is a symmetric matrix, i.e., U t = U .
It is also useful to define the right Cauchy–Green deformation matrix2 G = F t F .
When we apply the polar decomposition, we get
G = F t F = U t R t RU = U t U. (2.6)
This shows that G is independent of the rotation and is purely a function of the defor-
mation. In the case of a pure rotation, i.e., F = R, we find that G = I . This shows that
G in the case of a rotation is equal to identity. We also note that G has three scalar
invariants under a rigid coordinate transformation defined as follows:
1

I1 = trace(G), I2 =
trace(G)2 − trace G2 , I3 = det(G). (2.7)
2
In particular, in the case of an incompressible material det(G) = I3 = 1. For complete-
ness we also define the Green strain matrix E as E = 12 (G − I ). We next consider the
important case of small deformations and rotations.
Small deformations and rotations. If the deformations and the rotations are small (e.g.,
a maximal length change of the order of < 2–3%, and a maximum rotation of < 5◦ ), we
use the approximation (S PENCER [1980], Section 6.6)
∂u ∂u
≈ . (2.8)
∂x ∂X
From here we can rewrite F = RU as
F = RU = (I + ω)(I + ε). (2.9)

Here ω is the small rotation matrix and is antisymmetric. ε is the small (infinitesimal)
strain matrix and is symmetric. These are defined as

1 ∂u1 ∂u2 1
∂u1 ∂u3

0 2 ∂x2 − ∂x1 2 ∂x3 − ∂x1
1
 1
∂u2 ∂u1
∂u3 
ω = F − Ft =   2 ∂x − ∂x 0 1 ∂u2
2 ∂x − 
∂x2  ,
2
1 ∂u3
1
∂u1
2

1 ∂u3 ∂u2 3
2 ∂x1 − ∂x3 2 ∂x2 − ∂x3 0
2 In continuum mechanics literature this would be defined as the Cauchy–Green deformation tensor. A ma-
trix in that terminology is simply a two-dimensional tensor. In this chapter, we avoid the term tensor and use
the term matrix instead to improve general readability.
 ∂u1
1 ∂u1 ∂u2
1 ∂u1 ∂u3 
∂x 2 ∂x2 + ∂x1 2 ∂x3 + ∂x1
1
 1
∂u 1 ∂u1

∂u3 
ε= F +F t
−I =
 2 ∂x1 +
2
∂x2
∂u2
∂x2
1 ∂u2
2 ∂x3 + ∂x2  . (2.10)
2
1 ∂u3 ∂u1
1 ∂u3 ∂u2 ∂u3

2 ∂x1 + ∂x3 2 ∂x2 + ∂x3 ∂x3
Often, taking advantage of the symmetries these matrices are written in vector form as
e = [ε11 , ε22, ε33 , ε12 , ε13 , ε23]t , θ = [0, 0, 0, ω12, ω13 , ω23 ]t . (2.11)
This e is the classical definition for strain in infinitesimal linear elasticity (S PENCER
[1980]). Using x, y, z to represent the coordinate axes, e can also be written as
e = [εxx , εyy , εzz , εxy , εxz , εyz ]t . (2.12)

We note that the objectivity axiom is only approximately satisfied by the small defor-
mation approximation.
2.1.2. Material models

So far we have restricted our description to the geometry of the deformation. In this sec-
tion we extend this to account for what happens when a material deforms and relate the
deformation to the change in the internal structure of the material. Before proceeding to
give examples of possible material models, we first note that there are some theoretical
guidelines which must be observed (E RINGEN [1980]). The most important ones for
this work are:
(1) The axiom of objectivity – this requires the material model to be invariant with
respect to rigid motion or the spatial frame of reference.
(2) The axiom of material invariance – this implies certain symmetry conditions de-
pendent on the type of anisotropy of the material, and implicitly reduces the
number of free parameters.
The first axiom can be satisfied by postulating an internal or strain energy function W ,
which depends on the gradient deformation matrix F only through the Green defor-
mation matrix G, the Green strain matrix E, or in small deformation cases the infini-
tesimal strain matrix ε. The strain energy function serves as the material model. If we
postulate an internal energy which is not invariant to a global rotation, we arrive at the
following problem. Suppose that work is needed to rotate the object clockwise. From
conservation of energy principles, this energy will be returned when the object is turned
counter-clockwise. We can keep turning the object counter-clockwise to get more and
more energy and in this way we have created a perpetual motion machine and not a
material model.
Linear elastic energy functions. In this section e will be used to denote the vector
form of the infinitesimal strain matrix ε. The simplest useful continuum model in solid
mechanics is the linear elastic one. This is defined in terms of an internal energy function
W which has the form
W = et Ce, (2.13)
where C is a 6 × 6 matrix and defines the material properties of the deforming body.3
The simplest model is the isotropic linear elastic model used widely in the image analy-
sis literature. In this case the matrix C takes the form
 
1 −ν −ν 0 0 0
 −ν 1 −ν 0 0 0 
 
1  −ν −ν 1 0 0 0 
−1
C =   , (2.14)
E 0 0 0 2(1 + ν) 0 0 

 0 0 0 0 2(1 + ν) 0 
0 0 0 0 0 2(1 + ν)
where E is the Young’s modulus which is a measure of the stiffness of the material and
ν is the Poisson’s ratio which is a measure of incompressibility. This is the model that
will later be used to model brain deformation in the first case study of this chapter.
A transversely isotropic linear elastic model. For the second case study, involving the
left ventricle of the heart, we model the tissue using a transversely elastic material to
account for the preferential stiffness in the fiber direction. This is an extension of the
isotropic linear elastic model which allows for one of the three material axis to have a
different stiffness from the other two. In this case the matrix C takes the form
 1 −νp −νpf 
Ep Ep Ep 0 0 0
 −ν 
 p 1 −νpf 
 Ep Ep Ep 0 0 0 
 
 −νpf −νpf 1 
 0 0 0 
−1
C =  E p E p E f , (2.15)
2(1+νp ) 
 0 0 0 0 0 
 Ep 
 
 0 0 0 0 1
0 
 Gf 
1
0 0 0 0 0 Gf
where Ef is the fiber stiffness, Ep is cross-fiber stiffness and νpf , νp are the corre-
sponding Poisson’s ratios and Gf is the shear modulus across fibers (Gf ≈ Ef /(2(1 +
νfp )). If Ef = Ep and νp = νpf this model reduces to the more common isotropic lin-
ear elastic model. The fiber stiffness was set to be 3.5 times greater than the cross-fiber
stiffness (G UCCIONE and M C C ULLOCH [1991]). The Poisson’s ratios were both set
to 0.4 to model approximate incompressibility. The fiber orientations used are shown
in Fig. 2.2.
2.1.3. Stress and strain

While we have presented a material model formulation in terms of internal energy, an
alternative description of the material model is in terms of the stress–strain relationship
(S PENCER [1980]). The stress matrix can also be written in vector form (in the same
3 This class of model is linear as it results in a linear stress–strain relationship, i.e., σ = Ce. The term elastic
refers to the fact that the energy is completely recoverable, i.e., all energy used to compress the material is
returned once the compression forces are removed. Further, in this linear elastic formulation thermal effects
are ignored, which is equivalent to assuming constant temperature deformation.
F IG . 2.2. Fiber direction in the left ventricle as defined in G UCCIONE and M C C ULLOCH [1991]. Figure
reprinted from PAPADEMETRIS , S INUSAS , D IONE and D UNCAN [2001], Estimation of 3D left ventricular
deformation from echocardiography, Medical Image Analysis 5(1):17–29, ©2001 by permission from Else-
vier.
fashion as the strain matrix, see Eq. (2.12)) as

σ = [σxx , σyy , σzz , σxy , σxz , σyz ]t . (2.16)
If one pictures an infinitesimal cube of the material aligned with the coordinate axes
x, y, z, the normal stresses σxx , σyy and σzz are equal to the force per unit area acting
along the coordinate axes and on the faces of the cube. The shear stresses σxy , σxz and
σyz act along the surfaces of the cube.
Given an applied external stress σ , the material deforms resulting in material strain e.
If the material can be described using a linear elastic model, the stress–strain relation-
ship has the form
σ = Ce, (2.17)
where C is the 6 × 6 matrix from equation that models the material properties of the
object.
Next we present two possible frameworks for the integration of such material models
with image-derived information, the force equilibrium framework (Section 2.2) and the
energy minimization framework (Section 2.3). We note that the two are interchangeable
in the static case, but that the force equilibrium framework allows for the modeling of
material damping such as in visco-elastic materials.
2.2. The force equilibrium framework
In the force equilibrium framework the integration of image-derived information with

continuum mechanics material models is accomplished by converting the image-derived
measurements into boundary conditions for a resulting set of partial differential equa-
tions. First, we consider that the underlying material is in a state of static equilibrium,
which can be described in terms of the following equilibrium equations (H UEBNER ,
T HORNTON and B YROM [1995]):
∂σxx ∂σxy ∂σxz
+ + + Fx = 0,
∂x ∂y ∂z
∂σxy ∂σyy ∂σyz
+ + + Fy = 0, (2.18)
∂x ∂y ∂z
∂σxz ∂σyz ∂σzz
+ + + Fz = 0,
∂x ∂y ∂z
where F = (Fx , Fy , Fz ) are the externally applied forces.
To obtain the underlying displacement field we use the systems of Eqs. (2.17), (2.18)
and (2.10), as well as the definition of C from Eq. (2.14). By eliminating stress (σ ) and
strain (e) components, one can obtain:

1 ∂ ∂ux ∂uy ∂uz Fx
∇ 2 ux + + + + = 0,
1 − 2ν ∂x ∂x ∂y ∂z µ

1 ∂ ∂ux ∂uy ∂uz Fy
∇ uy +
2
+ + + = 0, (2.19)
1 − 2ν ∂y ∂x ∂y ∂z µ

1 ∂ ∂ux ∂uy ∂uz Fz
∇ 2 uz + + + + = 0,
1 − 2ν ∂z ∂x ∂y ∂z µ
where µ = E/(2(1 + ν)), and E, ν are the Young’s modulus and Poisson’s ratio, re-
spectively. These three equations are elliptic PDEs in displacements only and are known
as Navier equations (VALLIAPPAN [1981]). The image-derived information is used as
boundary conditions in the numerical solution of Eq. (2.19). This framework is ex-
ploited in case study I and an example solution is presented in Section 3.2.
By proper use of either a finite element or a finite difference discretization scheme,
Eq. (2.19) can also be rewritten in matrix form as
[K][U ] = [F ], (2.20)
where K is the global stiffness matrix, and U and F are the concatenated displacement
and force vectors, respectively. In particular, if the object is discretized to consist of n
nodes, the vector U has the form
U = [u1,x , u1,y , u1,z , u2,x , u2,y , u2,z , . . . , un,x , un,y , un,z ]t , (2.21)
where (up,x , up,y , up,z ) is the displacement of the pth node. The vector F similarly
consists of all the forces that act at each of the nodes. By appropriate manipulation of
the matrix K and the vector F , one can also impose displacement boundary conditions
as well (H UEBNER , T HORNTON and B YROM [1995]). We present an overview of the
finite element method in Section 2.5.
2.3. The energy minimization framework
In this section we describe a framework in which the goal is to estimate a displace-

ment field u which approximates another displacement field um . We will assume that
um is derived from some image-based algorithm, such as the shape-based tracking al-
gorithm, where the relationships between different displacements are not modeled. We
simplify the approximation problem to be a least-squares fit of u to um subject to some
constraints. This takes the form

2
û = arg min W (α, u, x) + c(x)um (x) − u(x) dv , (2.22)
u V
where:
• u(x) = (u1 , u2 , u3 ) is the vector valued displacement field defined in the region of
interest V and x is the position in space.
• um (x) = (um m m
1 , u2 , u3 ).
• c(x) is the spatially varying confidence in the measurements um .
• W (α, u, x) is a positive semi-definite functional which defines the approximation
strategy and is solely a function of u, a parameter vector α and the spatial posi-
tion x. In this work we will use the strain energy function (Eq. (2.13)) to set W .
This is commonly known as the regularization approach and W (α, u, x) is known as
the stabilization functional. In certain cases the input displacement field um is sparse
and is defined only on a finite number (P ) of points p within V . In this case the overall
functional takes the form

P
m 2
û = arg min W (α, u, x) dv + c(pi )u (pi ) − u(pi ) . (2.23)
u V i=1
Using principles from the calculus of variations, we can minimize the functionals de-
fined in Eqs. (2.22) and (2.23). In particular, using an appropriate discretization scheme,
the derivative form can be re-expressed in the same matrix notation as that of Eq. (2.20),
which then allows for the selection of an appropriate numerical solution scheme. We do
not discuss the details of the continuous case, instead we present an overview of the
finite element method in Section 2.5.
2.3.1. A probabilistic interpretation of the energy minimization framework

While we could simply substitute for the functional W in Eq. (2.23) with the appro-
priate internal energy function as defined in Eq. (2.13) (with C either from Eq. (2.14)
or (2.15)), we proceed to describe a probabilistic interpretation of the energy minimiza-
tion framework. The probabilistic interpretation is useful in indicating how one would
integrate noise-corrupted image derived data with a biomechanical model. We exploit
this framework in the second case study.
In the probabilistic interpretation of the energy minimization framework we again
aim to estimate the output displacements u from a set of measurements um . We further
assume that we are given the measurement probability density function p(um |u), which
also corresponds to the noise model for the measurements, and the prior probability den-
sity function for u, p(u).4 We pose this as a Bayesian a-posteriori estimation problem.
Within this framework, the solution û is the u that maximizes the posterior probability
density p(u|um ). Using Bayes’ rule, we can write the posterior probability as

p(u, um ) p(um |u)p(u)
û = arg max p u|um = = . (2.24)
u p(um ) p(um )
First, we note that p(um ) is a constant once the measurements have been made and can
therefore be ignored in the maximization process. We can rewrite the above expression
by taking logarithms to arrive at

û = arg max log p(u) + log p um |u . (2.25)
u
This expression is now in the same general form as Eq. (2.22). As previously demon-
strated by D. G EMAN and S. G EMAN [1984] and applied to medical image analysis
problems (e.g., C HRISTENSEN , R ABBITT and M ILLER [1994], G EE , H AYNOR , B RI -
QUER and BAJCSY [1997]), there is a correspondence between an internal energy func-
tion and a Gibbs probability density function. Given an energy function W (α, u, x) (not-
ing again that this can be expressed using the strain energy function; see Eq. (2.13)), we
can write an equivalent prior probability density function p(u) (see Eq. (2.24)) of the
Gibbs form (D. G EMAN and S. G EMAN [1984]):

p(u) = k1 exp −W (α, u, x) ,

log p(u) = log(k1 ) − W (α, u, x), (2.26)
where k1 is a normalization constant.
Next, we define the noise n = u − um . Then we can model the noise probabilistically,
using a multivariate Gaussian distribution, as
t −1
−n Σ n
p(n) = k2 exp ,
2
1
log p(n) = log k2 − nt Σ −1 n, (2.27)
2
where k2 is also a normalization constant and Σ is the covariance matrix which in this
case can be assumed to be diagonal, with the simplistic assumption that the noise is
uncorrelated. The mean of the noise is assumed to be equal to zero. Substituting for n
in this expression, we get

1
t

log p um |u = k2 − um − u Σ −1 um − u . (2.28)
2
By an appropriate choice of Σ, the second term can be mapped to the data adherence
term of Eq. (2.23). In this case Σ −1 will be a diagonal matrix with values c(pi ) on the
leading diagonal.
4 We will not define the basic terms of probability here, they can be found in standard textbooks such as
PAPOULIS [1991].
Advantages of the probabilistic interpretation. In the soft tissue deformation problem

there are usually two types of information: (i) the image derived data which is corrupted
by noise and (ii) the material properties of the soft tissue.
The data term is best modeled probabilistically in order to allow for the construc-
tion of a proper noise model. Here we can use ideas from the field of Digital Signal
Processing (see, for example, O PPENHEIM and S CHAFER [1975]). The material term,
however, is best defined in terms of a continuum mechanical model. The ability to gen-
erate an equivalent probability density function for an internal energy function, as was
done in Eq. (2.26), allows us to take a continuum mechanics model defined in terms
of an internal or strain energy function W (α, u, x), and generate a probability density
function p(u) which can then be used together with the probabilistic noise model within
a Bayesian estimation framework.
2.3.2. Soft tissue objects as Markov random fields

In using the Gibbs form (Eq. (2.26)), we have modeled the displacement field of the
solid probabilistically as a Markov random field, an example of this is shown in Fig. 2.3.
The Markov Random Field (MRF) then can be thought of as the probabilistic analog
of the continuum mechanical model. There are two interesting similarities: (i) both can
be defined using energy functions and (ii) the energy functions at any given point are
functions only of the values of that point and its immediate neighbors. In the case of the
MRF point (ii) comes from the fact that the Gibbs probability density function is often
defined on first and/or second order cliques which are very local neighborhoods of the
point. So if the displacement field is modeled as a MRF, the probability of the displace-
ment of a given point p effectively only depends on the displacement of its neighbors.
F IG . 2.3. Example of an object discretized by particles shown as black circles. If the displacement field is
modeled as a first-order Markov Random Field (MRF) the displacement of a specific particle p depends only
on external data and the displacements of its immediate neighbors a, b, c, d.
In the case of the mechanical model described using a strain energy function, the value
of the internal energy function, which via exponentiation in Eq. (2.26) becomes the
probability density function, at a given point depends only on the local strains. These
local strains are only dependent on the displacements of the neighbors of the point and
not on the displacements of the whole volume.
2.4. The active linear elastic model
The classical linear elastic model described in Eq. (2.13) is a passive model. In the
absence of any external force, the material will do nothing. Given no external work,
equilibrium is reached at the lowest energy state where the strain vector is identically
equal to zero. Such a material model is not accurate in the case of actively deforming
objects such as the left ventricle of the heart. In this case, a substantial part of the defor-
mation is actively generated by the muscle and is clearly not a result of external forces.
This active deformation does not produce a change in the strain energy of the material
and to account for this factor, we need to modify the elastic model appropriately. With
this in mind, we propose the active elastic model which takes the form
1
t

W= e − ea C e − ea , (2.29)
2
where ea is the active strain component. The active strain component represents the
deformation that is not a product of external forces and hence should not be penalized
by the model. In the absence of external forces, the active elastic model results in a
deformation equal to the one actively generated by the object. So in this sense it can
deform itself and hence it justifies the label active. Given a prior model of the active
contraction, the active elastic model can also be used to generate a prediction of the
position of the deforming object.
As an aside this model is also appropriate in the case where it is used to regularize an
image registration problem where there is no such physical notion of active deformation.
Here, the active component ea can be thought of as the expected magnitude of the
deformation.
Using the probabilistic interpretation to understand the active elastic model. By re-
ferring back to Section 2.3 and in particular to the expression of the internal energy
function as a Gibbs prior (see Eq. (2.26)), we can proceed to understand the rationale
for the active elastic model mathematically as follows. We first substitute for the inter-
nal energy functions of the active and the passive elastic models into Eq. (2.26). This
results in prior probability distributions of the form
−et Ce
Passive: log p(u) = k1 + , (2.30)
2
−(e − ea )t C(e − ea )
Active: log p(u) = k2 + , (2.31)
2
where k1 and k2 are normalization constants.
Note further that the standard multivariate normal distribution (mean = µ, covari-
ance = Σ) has the form (k3 is similarly a normalization constant):
−(u − µ)t Σ −1 (u − µ)
log p(u) = k3 + . (2.32)
2
By comparing Eqs. (2.30) and (2.31) to Eq. (2.32), we can see that in both cases
the material matrix C plays a similar role to the inverse of the covariance matrix (the
stiffer the material is, the greater the coupling between the displacements of neighboring
points and hence the smaller the effective component of the covariance matrix), and that
in the case of the active model, the active strain ea acts like the mean of the distribution.
In the case of the passive model, the mean is effectively zero. Hence, we can explicitly
see that the active elastic model is a generalization of the passive model, by adding the
possibility of having a non-zero mean. This is important in describing materials such as
the actively contracting tissue of the left ventricle.
2.5. The finite element method
The finite element method is a numerical analysis technique for obtaining approximate
solutions to a wide variety of engineering problems (H UEBNER , T HORNTON and B Y-
ROM [1995]). The key to this method is that the domain of problem is divided into
small areas or volumes called elements. The problem is then discretized on an element
by element basis and the resulting equations assembled to form the global solution.
2.5.1. An example problem

In this section we will describe an example problem and outline how it could be solved
using the finite element method. We will pose the problem in terms of an energy min-
imization framework where the goal is to estimate the displacement field u(x, y, z)
which is an optimal trade off between an internal energy function5 W (C, u) and ap-
proximating a noisy displacement field um (x, y, z) in a weighted least squares sense.
We define the optimal solution displacement field u is the one that minimizes func-
tional P (u). This is defined as

P (u) = W (C, u) + V u, um d(vol),
vol

2
W (C, u) = e(u)t Ce(u), V u, um = α um − u ,
where W (C, u) is the internal energy function defined by a strain energy function. C is
the constitutive law and e is the local strain which is a function of the displacements u.
V (u, um ) is the external energy term. um is the original (shape-tracking) displacement
estimate and α is the confidence in the match.
5 Note that although W is defined as function of the strain e, as e is a function of the displacement u, W can
also be written as a function of the displacement field u.
2.5.2. Outline of the solution procedure

Step 1. Divide volume into elements (tetrahedra or hexahedra) to provide the basis
functions for the discretization. In Fig. 2.4 a myocardium is shown tessellated into hexa-
hedral elements.
Step 2. Discretize the problem by approximating the displacement field in each el-
ement as a linear combinations of displacements at the nodes of each element. For a
hexahedral element this discretization can be expressed as

8
u≈ Ni ui ,
i=1
where Ni is the interpolation shape function for node i and ui is the displacement at
node i of the element. For the isoparametric hexahedral element shown in Fig. 2.5, we
define a local coordinate system ξi , and in this the shape functions Ni take the form
(H UEBNER , T HORNTON and B YROM [1995], Section 5.5):
1
Ni (ξ1 , ξ2 , ξ3 ) = (1 + ξ1 ξ1,i )(1 + ξ2 ξ2,i )(1 + ξ3 ξ3,i ), (2.33)
8
where (ξ1,i , ξ2,i , ξ3,i ) are the local coordinates of node i. It is easy to verify that the
shape function Ni takes a value of 1 at node i, a value of 18 at the origin and a value of 0
at all other nodes.
Step 3. Write down internal energy equation as the sum of the internal energy for
each element:

W (u) = et Ce d(vel ) . (2.34)
all elements vel
F IG . 2.4. A 3D hexahedral mesh generated by interpolating and filling between the endocardial and epicar-
dial boundaries. Figure reprinted from PAPADEMETRIS , S INUSAS , D IONE and D UNCAN [2001], Estimation
of 3D left ventricular deformation from echocardiography, Medical Image Analysis 5(1):17–29, ©2001 by
permission from Elsevier.
F IG . 2.5. Definition of local element coordinate system ξi and node coordinates for the nodes of a 2D 4-node
isoparametric element (left) and a 3D 8-node isoparametric element (right). For example, in the 2D case,
node 1 has coordinates (−1, −1). The centroid of the element O is the origin of the element specific coordinate
system. Note also that the axes are not necessarily orthogonal.
We further note that in an element we can approximate the derivatives of u with respect
to components of the global coordinate system x as follows (note that the ui are constant
in this expression):
∂u ∂(Ni ui ) ∂Ni
8 8
= = ui .
∂xk ∂xk ∂xk
i=1 i=1
However the shape functions Ni are expressed in terms of the local coordinate sys-
tem ξ . Using the chain rule, we can write
     
 ∂N  ∂x1 ∂x2 ∂x3  ∂N 

 
 
 

  
∂ξ1  ∂ξ1 ∂ξ1 ∂ξ1   ∂x1 


 ∂N      




 ∂x1 ∂x2 ∂x3  ∂N
= ×
 ∂ξ2 ∂ξ2 ∂ξ2   ∂x  (2.35)

 ∂ξ2 
    2

 
  ∂x1 ∂x2 ∂x3    


 ∂N    ∂N 
 
   
∂ξ3 ∂ξ3 ∂ξ3 ∂ξ3 ∂x3
or equivalently in matrix notation as Nξ = [J ] × Nx .
Hence we can calculate the desired derivatives Nx from the known derivatives Nξ by
inverting the Jacobian as follows: Nx = [J ]−1 Nξ . As long as the elements do not have
intersecting sides the Jacobian will remain invertible.
Note also that the derivatives of the displacement field u (i.e., ∂u/∂xk ) are a linear
function of the nodal displacements ui . Since the infinitesimal strain tensor consists of
only sums and differences of partial derivatives (see Eq. (2.10)) the infinitesimal strain
tensor can also be expressed as a linear function of the nodal displacements.6 This can
6 The finite strain deformation case is non-linear and does not allow for this simplification. The subsequent
expressions are so complicated that it makes the material beyond the scope of this brief overview. The reader
is referred to BATHE [1982].
be written in matrix form as e = Bu. Substituting this in Eq. (2.34), we get

W (u) = U et B t CB d(vel ) U e = U et K e U e,
all elements vel all elements
where K e7 is the element stiffness matrix,8 and U e is a vector obtained by concatenating

all the displacements of the nodes of the element, i.e.,
U e = [u1,x , u1,y , u1,z , . . . , u8,x , u8,y , u8,z ],
where ui = (ui,x , ui,y , ui,z ) is the displacement of node i.
Step 4. Rewrite the internal energy function in matrix form. First, we define the global
displacement vector U as
U = [u1,x , u1,y , u1,z , u2,x , u2,y , u2,z , . . . , un,x , un,y , un,z ]t , (2.36)
where n is the total number of nodes for the solid. We also define the global stiffness
matrix K as the assembly of all the local element stiffness matrices K e as

K= I Ke , (2.37)
all elements
where I is the re-indexing function. This takes an element Kije and adds it to the ele-
ment Kkl , where k and l are the global node numbers of local nodes i and j .9
The internal energy can now be written as W (U ) = U t KU .
Step 5. Write down the external energy function as a weighted least squares term,

n

2
V (u) = αi uei − ui .
i=1
If there is no initial displacement estimate for a given node j , set αj = 0.
7 The integration is carried out using Gaussian quadrature (H UEBNER , T HORNTON and B YROM [1995]).
8 Each component of K e indicates the ‘stiffness’ between any two nodes. One could in some sense think of
K14e as the stiffness of a spring connecting the x-directions of local nodes 1 and 2. (This ‘2’ is not a typo. The
first three rows of K e correspond to the components of the displacement of node 1, the second three to the
displacement of node 2, etc. See the definition of U e .)
9 Within an element the nodes are always numbered from 1 to 8. However this is a local index (short-
hand) to the global node numbers. When the global matrix is assembled the local indices (1 to 8) need to
be converted back to the global indices (e.g., 1 to n). K e has dimensions 24 × 24 and K has dimensions
3n × 3n. K14e , which is the stiffness between the x-directions of local nodes 1 and 2 would be part of K
kl
where k = 3(a − 1) + 1 and a is the global index of local node 1 and l = 3(b − 1) + 1, where b is the global
index of local node 2. Since nodes appear in more than one element the final value of Kkl is likely to be the
sum of a number of local Kije ’s.
Step 6. Rewrite external energy in a matrix form: we define the global initial displace-
ment vector U m in the same way as U above (see Eq. (2.36)) and the global confidence
matrix A to be a diagonal matrix with the confidence values for each displacement (αi )
forming the elements of the leading diagonal as follows:
 
a1
 a1 
 
 a1 
 
A=  ... . (2.38)

 a 
 n 
 an 
an
The external energy can be rewritten as V (U ) = (U m − U )t A(U m − U ).
Step 7. Form total potential energy equation P (U ) = W (U ) − V (U ).
Step 8. Solve for U. Differentiate P (U ) w.r.t. U and set to 0. This results in the final
equation

KU = A U m − U .
This is then solved for U using sparse matrix methods.10 U represents
! the values of u at
the nodes, and by means of the finite element approximation (u ≈ 8i=1 Ni ui ) we can
compute the resulting values of the displacement field u anywhere in the volume.
Having described the general common framework for the use of biomechanical mod-
els in the estimation of non-rigid displacement fields from medical images, we now
proceed to two specific case studies: (i) the estimation of brain shift for image guided
neurosurgery and (ii) the estimation of left ventricular deformation.
3. Case study I: brain shift compensation for image guided neurosurgery
3.1. Background
The use of surgical navigation systems has become a standard way to assist the neu-
rosurgeon in navigating within the intraoperative environment, planning and guiding
the surgery. One of the most important features of these systems is the ability to re-
late the position of the surgical instruments to the features in the preoperative images.
Ideally, they should provide a 3D display of the neuroanatomical structures of inter-
est and include visualization of surgical instruments within the same frame. In order
to be reliably used, the surgical navigation system should be as precise as possible,
preferably to within the voxel size of the dataset used (see G RIMSON , E TTINGER ,
W HITE , G LEASON , L OZANO -P EREZ , W ELLS III and K IKINIS [1996]). Most of the
10 In the case of finite deformations we end up with an expression of the form K(U ) = A(U m − U ) which
is solved iteratively.
current systems use preoperatively-acquired 3D data and register it to the patient coor-
dinate system (G RIMSON , E TTINGER , W HITE , G LEASON , L OZANO -P EREZ , W ELLS
III and K IKINIS [1995], G RIMSON , E TTINGER , W HITE , G LEASON , L OZANO -P EREZ ,
W ELLS III and K IKINIS [1996], P ETERS , DAVEY, M UNGER , C OMEAU , E VANS and
O LIVIER [1996], C HABRERIE , O ZLEN , NAKAJIMA , L EVENTON , ATSUMI , G RIM -
SON , K EEVE , H ELMERS , R IVIELLO , H OLMES , D UFFY, J OLESZ , K IKINIS and B LACK
[1998]). However, they assume that the brain and other intracranial structures are rigid
and fixed relative to the skull. The preoperative data is registered to the patient coor-
dinate system at the beginning of the surgery. While this can be done with a precision
to within 1 mm at start (G RIMSON , E TTINGER , W HITE , G LEASON , L OZANO -P EREZ ,
W ELLS III and K IKINIS [1996]), the brain deforms within the skull over time and thus
the accuracy of the system deteriorates. The median brain shift of points on the brain
surface has been estimated to range from 0.3 to 7.4 mm (H ILL , M AURER , WANG ,
M ACIUNAS , BARWISE and F ITZPATRICK [1997]). It is clear that a system based on the
rigid brain assumption cannot achieve a precision better than a few millimeters at the
outer structures. Since the deeper brain structures deform less than the outer ones the
error is the largest at the cortical surface. Obviously, the brain deforms even more after
interventions, e.g., post-resections. The average brain shift for cases in which hematoma
or tumors were removed has been reported to be even larger: 9.5 and 7.9 mm, respec-
tively (B UCHOLZ , Y EH , T ROBAUGH , M C D URMONT, S TURM , BAUMANN , H ENDER -
SON , L EVY and K ESSMAN [1997]). In our research, we are mainly concerned with
(but not limited to) issues surrounding epilepsy surgery where the amount of brain shift
of concern is more in line with that cited by H ILL , M AURER , WANG , M ACIUNAS ,
BARWISE and F ITZPATRICK [1997], although the physical implantation and removal of
subdural electrode grids further affect the amount of physical deformation.
Relatively little effort has been put forth to attempt to compensate for the defor-
mation that the brain undergoes during a surgical procedure. One line of investigation
incorporates the use of intraoperative MRI (iMRI) to periodically acquire full sets of
3D MR images that can be matched to preoperative anatomical datasets (G ERING ,
NABAVI , K IKINIS , et al. [1999], H ATA , NABAVI , WARFIELD , et al. [1999], H ILL ,
M AURER , M ARTIN , et al. [1999], F ERRANT, NABAVI , M ACQ , J OLESZ , K IKINIS
and WARFIELD [2001], NABAVI , B LACK , G ERING , et al. [2001], M IGA , ROBERTS ,
K ENNEDY, P LATENIKI , H ARTOV, L UNN and PAULSEN [2001], WARFIELD , TALOS ,
T EI , et al. [2002]). The cost-effectiveness and true utility of iMRI remains an open
question. Other attempts at this have used intraoperative ultrasound imaging (B U -
CHOLZ , Y EH , T ROBAUGH , M C D URMONT, S TURM , BAUMANN , H ENDERSON , L EVY
and K ESSMAN [1997]), physical modeling (M IGA , PAULSEN , K ENNEDY, H OOPES ,
H ARTOV and ROBERTS [1998], M IGA , ROBERTS , K ENNEDY, P LATENIKI , H ARTOV,
L UNN and PAULSEN [2001]), and includes our own work in using sparse sets of points
to update a physical model (S KRINJAR and D UNCAN [1999]), but all of these ideas re-
main in the very earliest stages of investigation and validation. There is additional work
in the field such as AUDETTE , S IDDIQI and P ETERS [1999].
3.2. System description
Our approach to brain shift compensation employs a 3D biomechanical brain model

(S KRINJAR , NABAVI and D UNCAN [2002]) guided by some limited interoperatively
acquired data. During the surgery we can use the model output to display preopera-
tive data (deformed according to the current model state). Before the surgery one can
acquire anatomical (MRI, CT) and functional (functional MR, SPECT, PET, . . .) im-
ages, segment them, generate surfaces of the segmented structures of interest, and then
deform all of them intraoperatively based on the current model state. If the model de-
formation prediction is close to the actual brain deformation, then the displayed images
and structures of interest (that are deformed according to the current model state) are
closer to the current actual brain state than they would be if one did not use the brain
shift compensation, making the surgical navigation system more precise and reliable.
An example of the effect of the brain shift is shown in Fig. 3.1. We note that the esti-
mation of interior displacements will be based on surface information produced from a
10–12 cm craniotomy that occurs during the first stage of these surgical procedures.
Therefore, we propose a biomechanical-model-based brain shift compensation sys-
tem composed of the following steps: preoperative image acquisition, segmentation,
mesh generation, registration of the model to the intraoperative environment, model
setup and guidance, and visualization of model-updated preoperative data.
3.2.1. Segmentation, visualization and registration

The first step after the preoperative image acquisition is the segmentation of the brain
tissue. For this task we have adopted an approach based on the automated algorithm
F IG . 3.1. Intraoperatively recorded points on the exposed brain surface at the beginning of the surgery are
shown at left, while their positions about 45 minutes later relative to the same pre-deformation brain surface
are shown at right. Gravity is perpendicular to the sagittal plane. The points moved in the direction of gravity
and they are hidden under the pre-deformation brain surface (only one of the points is still visible in the
figure at right). Since the brain deformed (in the direction of the gravity vector), the surface points moved
relative to the pre-deformation brain surface. Figure reprinted from S KRINJAR , N ABAVI and D UNCAN [2002],
Model-driven brain shift compensation, Medical Image Analysis, 6(4):361–373, ©2002 by permission from
Elsevier.
F IG . 3.2. A typical model mesh. The left figure shows the mesh, while the right one shows the mesh and
the outer brain surface. The mesh has over 2000 nodes and 1500 elements (bricks). Figure reprinted from
S KRINJAR , N ABAVI and D UNCAN [2002], Model-driven brain shift compensation, Medical Image Analysis,
6(4):361–373, ©2002 by permission from Elsevier.
suggested in S TOKKING [1998]. The main idea is to, after thresholding the brain MR
image (the threshold selection is the only manual input), a couple (typically three) bi-
nary erosions are performed to disconnect the brain tissue from the rest. After that, the
largest connected object, which is brain, is selected and then dilated the same number
of times as the thresholded image was eroded. The output of the dilations are masked
with the thresholded image to insure that the resulting object is within the brain tissue.
This simple and fast algorithm produces brain segmentation results of sufficient quality
for this project, since the mesh (which generation is based on the brain segmentation)
for the biomechanical model does not require the finest geometric details of the brain.
For object surface rendering we have used an improved version of the algorithm
presented in G IBSON [1998]. Some of the surfaces produced by this algorithm can be
seen in Fig. 3.2.
In order to display and use brain surface data for model guidance, a rigid body trans-
formation between the patient and preoperative image coordinate systems has to be
established. For this purpose we used a set of fiducial markers placed on the patient’s
skin. In the operating room (OR), the marker coordinates were recorded using a me-
chanical localizer (OMI [1997]). In addition, the markers were manually localized in
the preoperative MRI dataset.11 Then a robust point matching algorithm for resolving
the correspondences and finding the optimal rigid body transformation between the two
sets of marker locations was applied. The approach relies on the method for computing
11 Markers have to be visible in both MR and/or CT image data.
the optimal (in the least squares sense) rigid body transformation between two sets of
points with known correspondences (A RUN , H UANG and B LOSTEIN [1987]). To es-
tablish proper point correspondences, for the three most distant points in the first set,
all possible ordered triple of points from the second set are tested (note that there are
N(N − 1)(N − 2) triples, where N is the number of points; this is not computationally
demanding, since if N = 15, there are less than 3000 triples to check). For an ordered
triple of points from the second set, the optimal rigid body transformation is computed
from the three selected points in the first set to the triple. This rigid body transformation
is used to establish correspondences (based on the closest point criterion) for all the
points. Once the correspondences are known for all the points, the optimal rigid body
transformation is computed, and the sum of squared distances between corresponding
points is stored. This is repeated for all the ordered triples in the second set of points, and
the rigid body transformation that yields the smallest sum of squared distances is taken
as the final one. Once the rigid body transformation is determined, any point recorded
by the localizer can be mapped to the preoperative image coordinate system.
3.2.2. Mesh generation

The next step is to generate the model mesh from the segmented brain tissue. Here
we use hexahedral (“brick”) elements having 8 nodes at the vertex positions. The seg-
mented object (the brain tissue in this case) is the input to our mesh generator, which
generates an unstructured mesh (L ISEIKIN [1999]). The algorithm first generates a reg-
ular 3D matrix of bricks over the full 3D image. Each brick that has at least a half of its
volume inside the segmented object is kept, and others are discarded. The kept bricks
will compose the final mesh, while their nodes will be finely readjusted. The nodes are
divided into two groups. Each node that has all of its neighboring nodes left is called an
interior node, and all other nodes are called surface nodes. Each surface node is moved
to the closest point on the surface of the segmented object. Note that surface nodes be-
fore moving were not far from the surface of the segmented object. Finally, the interior
nodes are smoothed using a Laplacian-type smoother, in order to enhance the regularity
of the mesh. A typical output of the mesh generator is shown in Fig. 3.2. The meshes
we use do not capture all of the fine details of the segmentation output, but they still
achieve a reasonable performance in terms of accuracy and speed.
3.2.3. Image-based displacement estimates

There are different types of intraoperative data available for model guidance: points
(e.g., using a localizer and recording brain surface points over time), surface data: ob-
tained by a range system (AUDETTE , S IDDIQI , F ERRIE and P ETERS [2003]) or by a
stereo camera system (S KRINJAR , TAGARE and D UNCAN [2000]), and volumetric data
obtained by intraoperative image acquisition systems (intraoperative MRI, CT and ul-
trasound).
For this model we assume the use of a pair of stereo cameras overlooking the exposed
brain surface to acquire intraoperative information about the deforming brain. The idea
is to reconstruct and track the exposed brain surface as it deforms during the surgery. If
this can be done reliably, one can use the reconstructed brain surface as displacement
boundary conditions for the model PDEs. Each time the surgeon moves her or his hands
and surgical tools out of the way of the cameras, snapshots from the two cameras are
taken, exposed brain surface is reconstructed, the surface is used to guide the model, and
once the model is deformed, it can be used to update (properly warp) all preoperative
images available.
In these early results we simulate exposed brain surface tracking12 using intraoper-
ative MR images. In particular, we manually segmented the deformed brain from the
intraoperative scan and generated its surface. Since the brain surface did not move sig-
nificantly, we computed the displacement at each point r1 of the undeformed brain sur-
face S1 (only at the part of the brain surface that was visible through the craniotomy, i.e.,
at the exposed brain surface), as r = r2 − r1 , where r2 is the point on the deformed
brain surface S2 closest to the point r1 , i.e., obtained as argr2 ∈S2 min r2 − r1 . A more
advanced version of this approach is given in B ESL and M ACKAY [1992].
Finally, the computed displacements at the exposed brain surface were used as bound-
ary conditions for the partial differential equations derived using the model to be de-
scribed in the next section.
3.2.4. Mechanical model-based integration

We use a simple linear elastic model of brain deformation based on the following three
assumptions: (a) we desire a relatively simple model. Due to the complexity of the brain
shift phenomenon, not only that it is difficult to model some of the causative factors,
but also it is not clear how to set the model parameters (any increase in the model
complexity inevitably involves more parameters). Therefore, we base our approach on
a simple model, that incorporates the main tissue characteristics (elasticity and near-
incompressibility). The complexity of the deformation is made up by intraoperative
guidance of the model. (b) We assume a static model. Since intraoperative brain de-
formation is a relatively slow process with negligible dynamic components, we use a
static model. (c) We have some intraoperative input. The model has to by guided by
intraoperative input.
Brain shift is a small deformation relative to the brain size, and it is a good approx-
imation to use a linear stress strain relation and the infinitesimal strain approximation.
Although brain tissues are not isotropic, especially white matter due to its fibrous struc-
ture, since the fiber directions are not currently available to us, we assume that brain
tissues are isotropic materials.13 Due to the toughness of falx and tentorium, the move-
ment of the two structures is negligible in most cases. For this reason we fix the corre-
sponding parts of the model, i.e., we consider only the brain hemisphere on the side of
the craniotomy, and assume that the other brain parts do not deform.14
We integrate the material model and the image derived displacements using the force
equilibrium framework presented in Section 2.2, resulting in the following partial dif-
12 In a complete system brain surface tracking would be done by using a pair of stereo cameras.
13 Note however that the nodes in the model mesh that are located on the walls of the ventricles were set to
be free nodes. This allowed for “free” movement (the movement is constrained by the rest of the model) of
the walls of the ventricles. For example, this allows for relatively good modeling of the collapse of lateral
ventricles, which sometimes happens in brain surgery.
14 We note that the brain–skull interaction is not modeled directly. Rather, the hemisphere of the brain on
the side of the craniotomy is considered, the model nodes corresponding to falx and tentorium are fixed. This
indirectly models the effect of the skull on the brain at locations opposite to the craniotomy.
ferential equations:

1 ∂ ∂ux ∂uy ∂uz Fx
∇ 2 ux + + + + = 0,
1 − 2ν ∂x ∂x ∂y ∂z µ

1 ∂ ∂ux ∂uy ∂uz Fy
∇ 2 uy + + + + = 0,
1 − 2ν ∂y ∂x ∂y ∂z µ

1 ∂ ∂ux ∂uy ∂uz Fz
∇ 2 uz + + + + = 0.
1 − 2ν ∂z ∂x ∂y ∂z µ
These equations need to be solved with given displacement boundary conditions. Since
they are linear PDEs, and since differentiation is a linear operator, one can separately
find the solution u1 = (u1x , u1y , u1z ) for the equations with zero boundary conditions, and
the solution u2 = (u2x , u2y , u2z ) for the equations with zero body force, and the total solu-
tion will be u = u1 + u2 . However, gravity acts all the time, both before and during the
brain deformation, and therefore u1 will be the same in both cases. Since we are inter-
ested in the displacement field between the deformed and undeformed state, we do not
need to compute u1 . Thus, we need to solve only for u2 , i.e., solve Eq. (2.19) with the
given boundary conditions and zero body force. One should notice that gravity affects
u2 through boundary conditions (since the brain deforms partly because of gravity, and
a part of the brain surface will be used to define the displacement boundary conditions
– there are no explicit force boundary conditions in this case).
Another interesting observation is that Young’s modulus does not affect the displace-
ment field (u2 ), since the body force is zero in this case, and therefore the last terms in
Eq. (2.19) containing E (hidden in µ) disappear. Thus, the only model parameter to be
set is Poisson’s ratio. We have tested a range of values for ν, and the one that yielded
the smallest error (a partial validation is presented in Section 3.3) was ν = 0.4, which is
a value used by other groups as well (F ERRANT, WARFIELD and NABAVI [2000]). We
assume that the model is homogeneous since there is no reliable way known to us for
setting the model parameter for different brain structures.
3.3. Experimental results
In this section we will not present a complete stereo-guided brain deformation compen-
sation system, but rather we will investigate how well a continuum mechanics-based
brain model can predict in-volume deformation using only partial (exposed brain) sur-
face data for model guidance. We test our method using intraoperative MR image se-
quences. We segment the brain and construct a mesh composed of hexahedral (“brick”)
elements (with 5 mm approximate side lengths) was generated using the segmented data
and the in-house mesh generator described earlier. The generated meshes (of the cere-
bral hemisphere at the side of the craniotomy) had about 6500 nodes and about 5000
“brick” elements. Here we used the anatomical constraints that the falx and tentorium
are practically fixed, and we fixed the corresponding model nodes. For this reason it is
enough to consider only the half of the brain at the side of the craniotomy, since the
other part does not deform. We are aware that, although this assumption holds in most
cases, there are exceptions where falx moved during the surgery.
We present here a partial validation of the method using intraoperative MRI for two
cases:15 a sinking brain and a bulging brain. For both cases we generated the model and
displacement boundary conditions as explained above. We used ABAQUS (H IBBITT,
K ARLSSON and S ORENSEN [1997]) to compute the model deformation. For a model of
about 6500 nodes and about 5000 “brick” elements, it took about 80 seconds to solve the
equations on an SGI Octane R12000 computer. This time is almost practically applica-
ble, since it would mean that after about minute and a half after obtaining exposed brain
surface data, one would get updated MR images and other preoperative data. In order
to validate the computed deformation, we manually selected a set of anatomical land-
marks16 in the preoperative scan of the (undeformed) brain at various locations through-
out the volume of the cerebral hemisphere at the side of craniotomy. Then we manually
found the corresponding landmarks in the intraoperative scan of the (deformed) brain.
Finally, using the displacement field computed by the model, we determined the posi-
tions of the “model predicted landmarks” in the deformed brain corresponding to the
landmarks in the undeformed brain, and compared them to the corresponding manually
set landmarks in the deformed brain.
One can see from Table 3.1 that the maximal true landmark displacement was 3.8 mm
(3.6 mm) while the maximal error was 1.4 mm (1.3 mm) for the case of the sinking
(bulging) brain. Fig. 3.3 shows an MR image slice of a preoperative brain, the cor-
responding intraoperative image slice of the deformed brain, and the corresponding
model-updated image slice of the deformed brain. The maximal deformation was at the
exposed brain surface (about 7 mm for both cases). However, we did not use landmarks
close to the exposed brain surface since the exposed brain surface displacement was
used as a boundary condition, and the error at such landmarks would be unrealistically
small. Rather, we selected landmarks throughout the volume of the cerebral hemisphere
at the side of craniotomy away from the exposed brain surface. This is why the maximal
landmark displacement was under 4 mm.
TABLE 3.1
Case I (sinking brain) and case II (bulging brain): true landmark displacements (t), computed landmark dis-
placements (c), and error between true and computed landmark locations (e = c − t ), for 14 landmarks. All
values are in millimeters
1 2 3 4 5 6 7 8 9 10 11 12 13 14
I t 0.7 0.9 0.6 0.1 2.3 2.9 2.1 1.0 1.9 2.7 0.8 0.8 2.1 3.8
c 0.3 0.5 0.7 0.2 1.7 2.4 1.4 0.7 1.3 1.8 0.4 0.5 1.9 3.0
e 0.8 1.4 0.4 0.2 0.7 1.3 1.4 0.4 1.2 1.3 0.4 0.8 1.0 1.2
II t 2.7 1.8 0.6 3.6 2.6 0.8 1.3 1.1 1.4 0.7 0.7 0.4 2.4 0.5
c 2.0 1.6 1.1 2.4 2.6 0.5 0.8 1.2 1.5 0.8 0.5 0.2 2.0 0.3
e 0.8 1.0 0.6 1.3 0.8 0.4 0.9 0.8 0.9 0.5 0.7 0.5 1.2 0.7
15 In both cases we used intraoperative MR images after the dura was opened and brain deformed, but before
any major resection occurred.
16 For landmarks, we used points at anatomical structures that can relatively easily be identified in both
preoperative and intraoperative images.
F IG . 3.3. (a) A preoperative coronal slice of a sinking brain, (b) the corresponding intraoperative slice of the
deformed brain, (c) the corresponding model-computed slice of the deformed brain. Axial slices (d), (e) and
(f) correspond to the bulging brain case (undeformed, deformed and model-computed, respectively). Note that
in both cases the exposed brain surfaced in the computed slice moved similarly as the corresponding surface
in the intraoperative slice. Figure reprinted from S KRINJAR , N ABAVI and D UNCAN [2002], Model-driven
brain shift compensation, Medical Image Analysis, 6(4):361–373, ©2002 by permission from Elsevier.
4. Case study II: estimation of 3D left ventricular deformation
4.1. Background
Acute coronary artery occlusion results in myocardial injury, which will progress from
the endocardium to the epicardium of the heart wall in a wavefront fashion. A pri-
mary goal in the treatment of patients presenting with acute myocardial infarction is to
reestablish coronary flow, and to interrupt the progression of injury, thereby salvaging
myocardium. Unfortunately, there are no universally accepted non-invasive imaging
approaches for the accurate determination of the extent of injury. Using conventional
measures of regional myocardial function, the extent of myocardial infarction is over-
estimated. This can be attributed to persistent post-ischemic dysfunction (“stunning”),
persistent myocardial hypoperfusion (“hibernation”) or mechanical tethering of normal
areas by the adjacent injured myocardium. This tethering can be seen at the lateral mar-
gins of an infarct, resulting in a viable although dysfunctional border zone. Motion of
the viable epicardium can also be constrained by injury of the underlying endocardial
myocardial tissue. The location and ultimate transmural extent of the injury has impor-
tant implications for long term prognosis of patients following myocardial infarction.
Those patients with transmural myocardial infarction are likely to dilate their left ven-
tricles over time, a condition termed left ventricular “remodeling”. The occurrence of
post-infarction remodeling carries a much worse long-term prognosis.
A number of laboratories have shown that a comprehensive quantitative analysis of
myocardial strain can more accurately identify ischemic injury than simple analysis
of endocardial wall motion or radial thickening (A ZHARI , W EISS , ROGERS , S IU and
S HAPIRO [1995]). Furthermore, the characterization of segmental strain components
has shown great promise for defining the mechanical mechanisms of tethering or remod-
eling (K RAMER , ROGERS , T HEOBALD , P OWER , P ETRUOLO and R EICHEK [1996],
M ARCUS , G OTTE , ROSSUM , K UIJER , H EETHAAR , A XEL and V ISSER [1997]). Ex-
perimental animal studies demonstrate that decreased circumferential shortening in my-
ocardial regions adjacent to the infarct zone relative to remote regions is associated with
late left ventricular remodeling (K RAMER , et al. [1993]). At present, quantitative non-
invasive measurement of 3D strain properties from images has been limited to special
forms of magnetic resonance (MR) acquisitions, specifically MR tagging and restricted
to mostly research settings.
The MR tagging approach to the measurement of myocardial strain was originally
developed, and then vigorously pursued further by two groups, one at the University of
Pennsylvania (A XEL [1998]) and the other at Johns Hopkins (M C V EIGH [1998]). In
general, there are three different approaches to estimating displacement data from MR
tagging. The first approach involves tagging in multiple intersecting planes at the same
time, and using the tag intersections as tokens for tracking (e.g., A MINI , C HEN , C UR -
WEN , M ANU and S UN [1998], K ERWIN and P RINCE [1998], Y OUNG , K RAITCHMAN ,
D OUGHERTY and A XEL [1995]). The second approach involves tagging in multiple in-
tersecting planes, one set of parallel planes at a time. Then, each tagging plane is used
separately to estimate the normal direction of motion perpendicular to the plane. This
generates a set of partial displacements (i.e., the component parallel to the tag lines
is missing) to be combined later (e.g., H ABER , M ETAXAS and A XEL [1998], D EN -
NEY J R . and P RINCE [1995]). The final approach uses a lower resolution modulation
technique and attempts to model the tag fading over time using the Bloch equations.
The displacements are then extracted using a variable brightness optical flow technique
(e.g., P RINCE and M C V EIGH [1992], G UPTA and P RINCE [1995]). The reader is also
referred to a recently published book (A MINI and P RINCE [2001]).
As an alternative to MR tagging, several investigators have employed changes in
phase due to motion of tissue within a fixed voxel or volume of interest to assist in
estimating instantaneous, localized velocities and ultimately cardiac motion and defor-
mation. While the basic ideas were first suggested by VAN D IJK [1984] and NAYLER ,
F IRMIN and L ONGMORE [1986], it was Pelc and his team (P ELC , H ERFKENS , S HI -
MAKAWA and E NZMANN [1991], P ELC [1991], N.J. P ELC , H ERFKENS and L. P ELC
[1992]) that first bridged the technique to conventional cine MR imaging and permitted
the tracking of myocardial motion throughout the cardiac cycle. This technique basi-
cally relies on the fact that a uniform motion of tissue in the presence of a magnetic
field gradient produces a change in the MR signal phase that is proportional to ve-
locity. In general, two approaches have emerged to assemble deformation information
from phase contrast images: (i) processing the data directly to estimate strain rate (e.g.,
W EDEEN [1992], P ELC [1991]) and (ii) integrating the velocities over time, via some
form of tracking mechanism to estimate displacements (e.g., M EYER , C ONSTABLE , S I -
NUSAS and D UNCAN [1996], C ONSTABLE , R ATH , S INUSAS and G ORE [1994], Z HU ,
D RANGOVA and P ELC [1997], H ERFKENS , N. P ELC , L. P ELC and S AYRE [1991]).
The use of computer vision-based techniques to estimate displacement is also pos-
sible. One approach to establishing correspondence is to track shape-related features
on the LV over time as reported by K AMBHAMETTU and G OLDGOF [1994], C OHEN ,
AYACHE and S ULGER [1992]), A MINI and D UNCAN [1992], M C E ACHEN , OWEN and
D UNCAN [1997] and S HI , S INUSAS , C ONSTABLE , R ITMAN and D UNCAN [2000].
This is the basis for much of our own work and is expanded later. In general, here
preliminary displacements are estimated by matching local curvatures from segmented
surfaces from consecutive time frames and then the estimates are smoothed to produce
final displacement values. We note that such methods were applied to modalities other
than magnetic resonance such as X-ray CT (S HI , S INUSAS , C ONSTABLE , R ITMAN
and D UNCAN [2000], PAPADEMETRIS , S INUSAS , D IONE , C ONSTABLE and D UNCAN
[2002]) and ultrasound (PAPADEMETRIS , S INUSAS , D IONE and D UNCAN [2001]).
Finally, some investigators have used the intensity of the images directly to track lo-
cal LV regions. S ONG and L EAHY [1991] used the intensity in ultrafast CT images to
calculate the displacement fields for a beating heart. In addition, other investigators have
used local image intensity or intensity-based image texture from echocardiographic im-
age sequences to track local positions over 2D image sequences (M AILLOUX , B LEAU ,
B ERTRAND and P ETITCLERC [1987], M EUNIER [1998]). These efforts, along with
some related MR tagging approaches (e.g., G UPTA and P RINCE [1995]) roughly fall
into the category of optical flow-based methods. With the exception of methods based
on magnetic resonance tagging and to a lesser extent MR phase contrast velocities, none
of the other methods is capable of estimating complete three-dimensional deformation
maps of the left ventricle.
4.2. System description
Following image acquisition, the images are segmented interactively. From the results of
the segmentation we construct a three-dimensional finite element representation of the
left ventricle and also estimate initial surface correspondences using the shape-tracking
approach. A dense motion field is then estimated using a transversely isotropic, linear-
elastic model, which accounts for the muscle fiber directions in the left ventricle. The
dense motion field is in turn used to calculate the deformation of the heart wall in terms
of strain in cardiac specific directions. We explore each of these steps in more detail
next.
4.2.1. Image acquisition

MR imaging was performed on a GE Signa 1.5 Tesla scanner with version 4.8 soft-
ware using the head coil (26 cm diameter) for transmission and reception. Short axis
images, such as those shown in Fig. 4.1, through the left ventricle were obtained
with the gradient echo cine technique using the following parameters: TE = 6 msec,
F IG . 4.1. Series of MR short-axis images from End Diastole (ED) to End Systole (ES).
TR = 40 msec, flip angle = 30◦ , 16 phases collected, 5 mm slices, matrix 256 × 256,
2 averages, FOV = 40 cm. A total of 16 contiguous 5 mm thick slices were collected,
by acquiring four sets of staggered short axis slices (4 slices/set) with a separation gap
of 20 and 5 mm offset. This sequence provides images with an in-plane resolution of
1.64 ×1.64 mm for a 256 ×256 matrix and a 5 mm resolution perpendicular to the imag-
ing plane. This sequence also provides excellent temporal resolution (16 frames/cardiac
cycle, ≈ 40 msec/frame).
4.2.2. Segmentation and surface reconstruction

The left ventricle is segmented on a slice by slice basis using a custom designed soft-
ware platform (PAPADEMETRIS , R AMBO , D IONE , S INUSAS and D UNCAN [1998]).
The segmentation algorithm results in a set of planar contours parameterized using b-
splines which are subsequently sampled to generate a discrete set of points on each
plane.
From these contours we reconstruct the endo- and epicardial surfaces in a two-step
procedure as follows: (i) we interpolate between contours to generate in-between con-
tours at the desired sampling distance. This results in an iso-sampled set of points in
three dimensions. (ii) We construct a surface mesh by forming triangles between the
points. The procedure is illustrated in Fig. 4.2.
4.2.3. Mesh generation

We proceed to describe the mesh-generation method used for generating a volumet-
ric model for the left ventricle, in terms of hexahedral elements (PAPADEMETRIS , S I -
NUSAS , D IONE , C ONSTABLE and D UNCAN [2002]). The output mesh of this algorithm
will be used to describe the geometry of the left ventricle as needed for the estimation
of the complete deformation field using finite elements. Here, we describe an algorithm
that takes advantage of the ‘cylinder-like’ geometry of the left ventricle to make the
problem easier. We first interpolate on a contour-by-contour basis between the endocar-
dial and epicardial surfaces using shape-based interpolation to generate an appropriate
number of in-between interpolated surfaces (typically 3 or 4). Because of the greater
geometrical complexity of the endocardium, we space the interpolated surfaces to be
preferentially closer to the endocardium. We then discretize the contour on the middle
slice of the endocardium to the desired number of nodes (typically 35–45). Then we es-
F IG . 4.2. Steps involved in moving from slice by slice contours to full surface representation.
(1) Slice-by-slice b-spline parameterized contours as extracted by the segmentation process. (2) Discretized
contours as equally-spaced points. (3) Formation of wire-frame by Delaunay triangulation. (4) Surface render-
ing. (5) Smoothing of surface using non-shrinking smoothing algorithm. (6) + (7) First and second principal
curvatures of the surface. Here, green shows negative (i.e., inward) curvature, white shows flat regions and
red indicates positive (i.e., outward) curvature. Figure reprinted from PAPADEMETRIS , S INUSAS , D IONE ,
C ONSTABLE and D UNCAN [2002], Estimation of 3D left ventricular deformation from medical images using
biomechanical models, IEEE Transactions on Medical Imaging, 21(7):786–800, ©2002 by permission from
the IEEE.
timate correspondences between the surfaces and connect corresponding points to form
hexahedral brick like elements.
4.2.4. Shape-based tracking

In this work, the original displacements on the outer surfaces of the myocardium were
obtained by using the shape-tracking algorithm whose details were presented in S HI , S I -
NUSAS , C ONSTABLE , R ITMAN and D UNCAN [2000]. The method tries to track points
F IG . 4.3. Example of shape-tracking approach. The goal here is to map the original surface to the final sur-
face. For a point p1 on the original surface a window W of plausible matching points on the final surface
is first generated. Then the point p2 in W which has the most similar shape-properties to p1 is selected as
the candidate match point. The distance function for shape-similarity is typically based on the curvature(s).
Figure reprinted from PAPADEMETRIS , S INUSAS , D IONE and D UNCAN [2001], Estimation of 3D left ven-
tricular deformation from echocardiography, Medical Image Analysis 5(1):17–29, ©2001 by permission from
Elsevier.
on successive surfaces using a shape similarity metric which tries to minimize the differ-
ence in principal curvatures and was validated using implanted markers (S HI , S INUSAS ,
C ONSTABLE , R ITMAN and D UNCAN [2000]).
With reference to Fig. 4.3, consider point p1 on a surface at time t1 which is to be
mapped to a point p2 on the deformed surface at time t2 . First, a search is performed a
physically plausible region W on the deformed surface, to find the point p̂2 which has
the local shape properties closest to those p1 . The shape properties here are captured in
terms of the principal curvatures κ1 and κ2 . The distance measure used is the bending
energy required to bend a curved plate or surface patch to a newly deformed state. This
is labeled as dbe and is defined as

(κ1 (p1 ) − κ1 (p2 ))2 + (κ2 (p1 ) − κ2 (p2 ))2
dbe (p1, p2) = A . (4.1)
2
The point p̂2 is found by minimizing dbe in the region W which can be expressed as

p̂2 = arg min dbe (p1, p2) . (4.2)
p2 ∈ W
Finally, displacement estimate vector for each point p1 , um

1 is given by
1 = p̂2 − p1 .
um
Confidence measures in the match. The bending energy measures for all the points in-
side the search region W are recorded as the basis to measure the goodness and unique-
ness of the matching choices. The value of the minimum bending energy in the search
region between the matched points indicates the goodness of the match. Denoting this
value as mg , we have the following measure for matching goodness:

mg (p1 ) = dbe (p1 , p̂2 ). (4.3)
On the other hand, it is desirable that the chosen matching point is a unique choice
among the candidate points within the search window. Ideally, the bending energy value
of the chosen point should be an outlier (much smaller value) compared to the values of
the rest of the points. If we denote the mean values of the bending energy measures of all
the points inside window W except the chosen point as d̄be and the standard deviation
d
as σbe , we define the uniqueness measure as
dbe (p1 , p̂2 )
mu (p1 ) = . (4.4)
d̄be − σbe
d
This uniqueness measure has a high value if the bending energy of the chosen point is
small compared to some smaller value (mean minus standard deviation) of the remain-
ing bending energy measures. Combining these two measures together, we arrive at one
confidence measure cm (p1 ) for the matched point p̂2 of point p1 :
1 1
cm (p1 ) = × , (4.5)
k1,g + k2,g mg (p1 ) k1,u + k2,u mu (p1 )
where k1,g , k2,g , k1,u and k2,u are scaling constants for normalization purposes. We
normalize the confidences to lie in the range 0 to 1.
Modeling the initial displacement estimates. Given a set of displacement vector mea-
surements um and confidence measures cm , we model these estimates probabilistically
by assuming that the noise in the individual measurements is normally distributed with
zero mean and a variance σ 2 = 1/cm . In addition, we assume that the measurements are
uncorrelated. Given these assumptions, we can write the measurement probability for
each point as

1 −(u−um )2
p um |u = √ e 2σ 2 . (4.6)
2πσ 2
4.2.5. Mechanical model-based integration

We model the left ventricle using the active elastic model described in Section 2.4. The
passive properties of the material (captured by the matrix C of Eq. (2.29)) are modeled
using the transversely isotropic elastic model described in Eq. (2.15), with preferen-
tial stiffness being applied along canonical fiber orientations. Such a fiber model was
shown in Fig. 2.2. The active elastic model results in a prior probability density function
for the underlying displacement field p(u) which is integrated with the image-derived
measurements p(um |u) (see Eq. (4.6)) to yield a maximum a-posteriori solution of the
form:

m p(um |u)p(u)
û = arg max p u|u = arg max . (4.7)
u u p(um )
The prior probability of the measurements p(um ) is a constant once these measurements
have been made and therefore drops out of the minimization process.
F IG . 4.4. Illustration of the two-step numerical solution technique.
Taking logarithms in Eq. (4.7) and differentiating with respect to the displacement
field u results in a system of partial differential equations, which we solve using the
finite element method (described in Section 2.5). The first step in the finite element
method is the division or tessellation of the body of interest into elements; these are
commonly tetrahedral or hexahedral in shape. Once this is done, the partial differential
equations are written down in integral form for each element, and then the integral of
these equations over all the elements is taken to produce the final set of equations. For
more information one is referred to standard textbooks such as BATHE [1982]. The final
set of equations is then solved to produce the output set of displacements. In our case
the myocardium is divided into approximately 2500 hexahedral elements.
For each frame between end-systole (ES) and end-diastole (ED), a two step problem
is posed: (i) solving Eq. (4.7) normally and (ii) adjusting the position of all points on
the endo-and epi-cardial surfaces so they lie on the endo- and epi-cardial surfaces at
the next frame using a modified nearest-neighbor technique and solving Eq. (4.7) once
more using this added constraint. This ensures that there is no bias in the estimation
of the radial strain. This is illustrated schematically in Fig. 4.4. Consider the point A
on the epicardial contour at time t. (The endo-cardial contours are not shown for the
sake of clarity.) After normal solution of Eq. (4.7) it gets mapped to point B which does
not lie on the epi-cardial contour at time t + 1. The point is then fixed to point C by the
modified nearest-neighbor technique and Eq. (4.7) is solved again to adjust the positions
of internal points.
4.3. Experimental results
In this section we present some preliminary results of the application of this algorithm to
left ventricular deformation estimation. We bootstrap the algorithm by using the output
produced by our previous work (PAPADEMETRIS , S INUSAS , D IONE , C ONSTABLE and
D UNCAN [2002]). We label this algorithm as the ‘passive’ algorithm. In the passive
algorithm, the images are segmented interactively and then initial correspondence is
established using a shape-tracking approach. A dense motion field is then estimated
using a passive, transversely linear elastic model, which accounts for the fiber directions
in the left ventricle. The dense motion field is in turn used to calculate the deformation
of the heart wall in terms of strains.
The output of the ‘passive’ algorithm consists of a set of vectors ep (xi , tj ) repre-
senting the strain estimated by the passive algorithm at position xi and time tj . Typ-
ically we divide the heart into about 800–1000 (i.e., i ∈ 1 : 1000) elements and use
6–9 time frames (j ∈ 1 : 9) resulting in a total of approximately 7000 6 × 1 vectors
p p p p p p
ep = [err , ecc , ell , erc , erl , elc ]t . The components of ep are the normal strains in the
radial (rr), circumferential (cc) and longitudinal (ll) directions as well as the shears
p
between these direction (e.g., erc is the radial-circumferential shear strain).
These vectors ep are then used to generate an estimate of the active strain ea using
isovolumic correction and possibly temporal smoothing. In the isovolumic correction
procedure at each discrete element position xi and time tj we generate an output vector
ea (xi , tj ) by adjusting the longitudinal strain to create a new set of strain estimates ea
that result in an incompressible deformation. These estimates ea are used as the mean
value for the active elastic model. The variance is determined by the stiffness matrix
and is the same as it was for the passive model. We label the results produced by this
procedure as Active.
Validation. We tested the new algorithm(s) by comparing its output to those ob-
tained using MR tagging (K ERWIN and P RINCE [1998]) and implanted markers (PA -
PADEMETRIS , S INUSAS , D IONE , C ONSTABLE and D UNCAN [2002]). In the MR tag-
ging case we used one human image sequence provided to us by Dr. Jerry Prince from
John Hopkins University. The images were acquired using 3 orthogonal MR tagging
acquisitions and the displacements estimated using an algorithm presented in K ERWIN
and P RINCE [1998]. From these displacements we estimate the MR tagging derived
strains. Images from one of the three acquisitions had the evidence of the tag lines re-
moved using morphological operators, was segmented interactively and the strains were
estimated using our previous approach (Passive) (PAPADEMETRIS , S INUSAS , D IONE ,
C ONSTABLE and D UNCAN [2002]). In the case of implanted markers we used 8 ca-
nine image sequences with implanted markers as is shown in Fig. 4.5 (see also PA -
PADEMETRIS , S INUSAS , D IONE , C ONSTABLE and D UNCAN [2002]).
We tested two permutations of the active algorithm. For the algorithm labeled Active
in Fig. 4.7, we used as input the output of the passive algorithm after isovolumic correc-
tion, without any temporal smoothing. The algorithm labeled as ActiveT used the out-
put of the passive algorithm with both temporal smoothing and isovolumic correction.
Fig. 4.6 illustrates the output of algorithm ActiveT at four points in the cardiac cycle as
applied to the MR tagging sequence. The output of the tagging method (K ERWIN and
P RINCE [1998]) at end-systole is presented for comparison.
Fig. 4.7 shows the error between the estimates of our old algorithm labeled passive
and the two variations of the new active algorithm (Active and ActiveT), as compared to
the output of the tagging algorithm (K ERWIN and P RINCE [1998]) and to the estimates
obtained using the MR markers. In the case of the tagging algorithm we observe an
overall reduction in mean strain error from 9.9% (passive) to 8.1% (active) at end-
F IG . 4.5. Implantation of Image-Opaque Markers. This figure shows the arrangement of markers on the
myocardium. First, a small bullet-shaped copper bead attached to an elastic string was inserted into the blood
pool through a needle track. Then the epicardial marker was sutured (stitched) to the myocardium and tied
to the elastic string. Finally, the mid-wall marker was inserted obliquely through a second needle track to
a position approximately half-way between the other two markers. Figure reprinted from PAPADEMETRIS ,
S INUSAS , D IONE , C ONSTABLE and D UNCAN [2002], Estimation of 3D left ventricular deformation from
medical images using biomechanical models, IEEE Transactions on Medical Imaging, 21(7):786–800, ©2002
by permission from the IEEE.
F IG . 4.6. Leftmost four columns: circumferential, radial and longitudinal strain outputs of our active (Active
2T) algorithm at four points in the systolic half of the cardiac cycle. Far right column: output of MR tagging
based algorithm (K ERWIN and P RINCE [1998]) on the same image sequence shown at the last time frame
(End-systole). Figure reprinted from PAPADEMETRIS , O NAT, S INUSAS , D IONE , C ONSTABLE and D UNCAN
[2001], The active elastic model, in: Information Processing in Medical Imaging, IPMI’01, Davis, CA, in:
Lecture Notes in Computer Science 2082, ©2001 by permission from Springer-Verlag.
systole (frame 10). In the case of the implanted markers we observe a similar reduction
from 7.2 to 6.3%.
It is also interesting to note that the MR tagging algorithm (K ERWIN and P RINCE
[1998]) produces a reduction of myocardial volume of 12% between end-diastole and
end-systole, our passive algorithm an increase of approximately 14% and all both ver-
F IG . 4.7. Absolute strain error vs tag data or implanter markers. Passive – passive model from PA -
PADEMETRIS , S INUSAS , D IONE , C ONSTABLE and D UNCAN [2002], Active and ActiveT represent two ver-
sions of the active algorithm without and with temporal smoothing. We note that both the active algorithms
result in error reduction as compared to the passive algorithm. In the case of the tagging data we plot the ab-
solute error in the cardiac-specific strains whereas in the case of implanted markers we use the principal strains
instead (see PAPADEMETRIS , S INUSAS , D IONE , C ONSTABLE and D UNCAN [2002]). Figure reprinted from
PAPADEMETRIS , O NAT, S INUSAS , D IONE , C ONSTABLE and D UNCAN [2001], The active elastic model, in:
Information Processing in Medical Imaging, IPMI’01, Davis, CA, in: Lecture Notes in Computer Science
2082, ©2001 by permission from Springer-Verlag.
sions of the active algorithm produced small increases (< 2%) showing that the isovo-
lumic correction was effective.
5. Conclusions
The use of biomechanical models to guide the estimation of non-rigid motion and de-
formation in medical image analysis is now well-established. Additional areas where
such models are used are in surgical simulation (C OTIN , D ELINGETTE and AYACHE
[1999]), intra-subject non-rigid brain registration before and after tumor growth (K YR -
IAKOU and DAVATZIKOS [1998]). Continuum mechanical models have also been used
purely for their mathematical properties in the case of the generic non-rigid registration
problem (e.g., C HRISTENSEN , R ABBITT and M ILLER [1996], G EE , H AYNOR , B RI -
QUER and BAJCSY [1997]).
In this chapter we have particularly focused on the integration of bio-mechanical
models with image-derived information and have presented two frameworks to achieve
this integration: (i) the force equilibrium framework and (ii) the energy minimization
framework. The force equilibrium framework results in a set of partial differential equa-
tions describing the displacement field and image-derived displacements are used as
boundary conditions. The energy minimization framework, which is further interpreted
and recast as a Bayesian estimation framework, similarly results in a set of partial differ-
ential equations which when discretized using the finite element method yield a system
of linear equations. With the use of an appropriate noise model the image-derived mea-
surements can be modeled probabilistically and integrated with the model which in turn
is also modeled as a Gibbs prior.
While some of the details of the extraction of image-derived measurements are pre-
sented, this is not the major point of the chapter, in this respect our work differs little
for other image analysis work utilizing mathematical regularization models such as in
P RINCE and M C V EIGH [1992]. The key issue with mathematical regularization models
is that the choice of parameters for the regularization functional is fairly arbitrary and
ad-hoc, the use of biomechanical models offers the advantage that the model parameters
are physically meaningful and can be experimentally measured, as is the case with the
fiber orientations in the left ventricle.
We expect that the sophistication of the models will continue to improve as re-
searchers begin to tackle more complex problems such as tumor resection. It is also
likely that more current sophisticated work in biomechanics such as that in the cardiac
modeling area (e.g., H UNTER , M C C ULLOCH and N IELSEN [1991]) will become more
relevant to state of the art medical image analysis. In this we are encouraged by the ap-
pearance of work such as that of S ERMESANT, C OUDIRE , D ELINGETTE and AYACHE
[2002] which is beginning the process of integrating not only the biomechanical prop-
erties of the left ventricle but also the electrical properties of the tissue within an image
analysis problem.
References
ABAQUS/Version 5.7 (1997). Hibbitt, Karlsson & Sorensen, Rhode Island, USA.
A MINI , A.A., C HEN , Y., C URWEN , R.W., M ANU , V., S UN , J. (1998). Coupled B-snake grids and con-
strained thin-plate splines for analysis of 2D tissue deformations from tagged MRI. IEEE Trans. Medical
Imaging 17 (3), 344–356.
A MINI , A.A., D UNCAN , J.S. (1992). Bending and stretching models for LV wall motion analysis from curves
and surfaces. Image and Vision Computing 10 (6), 418–430.
A MINI , A.A., P RINCE , J.L. (eds.) (2001). Measurement of Cardiac Deformations from MRI: Physical and
Mathematical Models (Kluwer Academic, Dordrecht).
A RUN , K.S., H UANG , T.S., B LOSTEIN , S.D. (1987). Least-squares fitting of two 3-d point sets. IEEE Trans.
Pattern Analysis and Machine Intelligence 9 (5), 698–700.
AUDETTE , M.A., S IDDIQI , K., F ERRIE , F.P., P ETERS , T.M. (2003). An integrated range-sensing, segmen-
tation and registration framework for the characterization of intra-surgical brain deformations in image-
guided surgery. Computer Vision and Image Understanding 89 (2–3), 226–251.
AUDETTE , M.A., S IDDIQI , K., P ETERS , T.M. (1999). Level-set surface segmentation and fast cortical range
image tracking for computing intrasurgical deformations. In: Medical Image Computing and Computer
Assisted Intervention, pp. 788–797.
A XEL , L. (1998). Physics and technology of cardiovascular MR imaging. Cardiology Clinics 16 (2), 125–
133.
A ZHARI , H., W EISS , J., ROGERS , W., S IU , C., S HAPIRO , E. (1995). A noninvasive comparative study of
myocardial strains in ischemic canine hearts using tagged MRI in 3D. Amer. J. Physiol. 268, H1918–
H1926.
BATHE , K. (1982). Finite Element Procedures in Engineering Analysis (Prentice Hall, Englewood Cliffs, NJ).
B ESL , P.J., M ACKAY , N.D. (1992). A method for registration of 3-D shapes. IEEE Trans. Pattern Analysis
and Machine Intelligence 14 (2), 239–256.
B UCHOLZ , R., Y EH , D., T ROBAUGH , J., M C D URMONT , L., S TURM , C., BAUMANN , C., H ENDERSON , J.,
L EVY , A., K ESSMAN , P. (1997). The correction of stereotactic inaccuracy caused by brain shift using an
intraoperative ultrasound device. In: CVRMed-MRCAS’97, Grenoble, France, pp. 459–466.
C HABRERIE , A., O ZLEN , F., N AKAJIMA , S., L EVENTON , M., ATSUMI , H., G RIMSON , E., K EEVE , E.,
H ELMERS , S., R IVIELLO , J., H OLMES , G., D UFFY , F., J OLESZ , F., K IKINIS , R., B LACK , P. (1998).
Three-dimensional reconstruction and surgical navigation in pediatric epilepsy surgery. In: Wells, W.,
Colchester, A., Delp, S. (eds.), MICCAI’98, pp. 74–83.
C HRISTENSEN , G.E., R ABBITT, R.D., M ILLER , M.I. (1994). 3D brain mapping using deformable neu-
roanatomy. Physics in Medicine and Biology 39, 609–618.
C HRISTENSEN , G.E., R ABBITT, R.D., M ILLER , M.I. (1996). Deformable templates using large deforma-
tion kinematics. IEEE Trans. Image Process. 5 (10), 1435–1447.
C OHEN , I., AYACHE , N., S ULGER , P. (1992). Tracking points on deformable objects using curvature infor-
mation. In: ECCV’92. In: Lecture Notes in Comput. Sci. (Springer-Verlag, Berlin), pp. 458–466.
C ONSTABLE , T., R ATH , K., S INUSAS , A., G ORE , J. (1994). Development and evaluation of tracking algo-
rithms for cardiac wall motion analysis using phase velocity MR imaging. Magn. Reson. Med. 32, 33–42.
C OTIN , S., D ELINGETTE , H., AYACHE , N. (1999). Real-time elastic deformations of soft tissues for surgery
simulation. IEEE Trans. Visualization and Computer Graphics 5 (1), 62–73.
D ENNEY J R ., T.S., P RINCE , J.L. (1995). Reconstruction of 3-D left ventricular motion from planar tagged
cardiac MR images: An estimation theoretic approach. IEEE Trans. Medical Imaging 14 (4), 625–635.
587
E RINGEN , A.C. (1980). Mechanics of Continua (Krieger, New York, NY).

F ERRANT , M., N ABAVI , A., M ACQ , B., J OLESZ , F., K IKINIS , R., WARFIELD , S. (2001). Registration of 3D
intraoperative MR images of the brain using a finite-element biomechanical model. IEEE Trans. Medical
Imaging 20 (12), 1384–1397.
F ERRANT , M., WARFIELD , S.K., N ABAVI , A.E. (2000). Registration of 3D intraoperative MR images of the
brain using a finite element biomechanical model. In: Medical Image Computing and Computer Assisted
Intervention, pp. 19–28.
G EE , J.C., H AYNOR , D.R., B RIQUER , L.L., BAJCSY , R.K. (1997). Advances in elastic matching theory
and its implementation. In: CVRMed-MRCAS, Grenoble, France.
G EMAN , D., G EMAN , S. (1984). Stochastic relaxation, Gibbs distribution and Bayesian restoration of im-
ages. IEEE Trans. Pattern Analysis and Machine Intelligence 6, 721–741.
G ERING , D.T., N ABAVI , A., K IKINIS , R., et al. (1999). An integrated visualization system for surgical
planning and guidance using image fusion and interventional imaging. In: Medical Image Computing and
Computer Assisted Intervention, MICCAI’99, Cambridge, UK, pp. 809–819.
G IBSON , S. (1998). Constrained elastic surface nets: Generating smooth surface from binary segmented data.
In: Medical Image Computing and Computer Aided Intervention, MICCAI’98, pp. 888–898.
G RIMSON , W.E.L., E TTINGER , G.J., W HITE , S.J., G LEASON , P.L., L OZANO -P EREZ , T., W ELLS III,
W.M., K IKINIS , R. (1995). Evaluating and validating an automated registration system for enhanced
reality visualization in surgery. In: CVRMed, Nice, France.
G RIMSON , W.E.L., E TTINGER , G.J., W HITE , S.J., G LEASON , P.L., L OZANO -P EREZ , T., W ELLS III,
W.M., K IKINIS , R. (1996). An automatic registration method for frameless stereotaxy, image guided
surgery, and enhanced reality visualization. IEEE Trans. Medical Imaging 15 (2), 129–140.
G UCCIONE , J.M., M C C ULLOCH , A.D. (1991). Finite element modeling of ventricular mechanics. In:
Hunter, P.J., McCulloch, A., Nielsen, P. (eds.), Theory of Heart (Springer-Verlag, Berlin), pp. 122–144.
G UPTA , S.N., P RINCE , J.L. (1995). On variable brightness optical flow for tagged MRI. In: Information
Processing in Medical Imaging.
H ABER , E., M ETAXAS , D.N., A XEL , L. (1998). Motion analysis of the right ventricle from MRI images. In:
Medical Image Computing and Computer Aided Intervention, MICCAI’98, Cambridge, MA, pp. 177–188.
H ATA , N., N ABAVI , A., WARFIELD , S., et al. (1999). A volumetric optical flow method for measurement
of brain deformation from intraoperative magnetic resonance images. In: Medical Image Computing and
Computer Assisted Intervention, MICCAI’99, Cambridge, UK, pp. 928–935.
H ERFKENS , R., P ELC , N., P ELC , L., S AYRE , J. (1991). Right ventricular strain measured by phase contrast
MRI. In: Proc. 10th Annual SMRM, San Francisco, p. 163.
H ILL , D., M AURER , C., WANG , M., M ACIUNAS , R., BARWISE , J., F ITZPATRICK , M. (1997). Estimation
of intraoperative brain surface movement. In: CVRMed-MRCAS’97, Grenoble, France, pp. 449–458.
H ILL , D.L.G., M AURER , C.R., M ARTIN , A.J., et al. (1999). Assessment of intraoperative brain deformation
using interventional MR imaging. In: Medical Image Computing and Computer Assisted Intervention,
MICCAI’99, Cambridge, UK, pp. 910–919.
H UEBNER , K.H., T HORNTON , E.A., B YROM , T.G. (1995). The Finite Element Method For Engineers (Wi-
ley, New York).
H UNTER , P.J., M C C ULLOCH , A., N IELSEN , P. (eds.) (1991). Theory of Heart (Springer-Verlag, Berlin).
H UNTER , P.J., N ASH , M.P., S ANDS , G.B. (1997). Computational electromechanics of the heart. In: Panfilov,
A.V., Holden, A.V. (eds.), Computational Biology of the Heart (Wiley, New York), pp. 346–407.
K AMBHAMETTU , C., G OLDGOF , D. (1994). Curvature-based approach to point correspondence recovery in
conformal nonrigid motion. CVGIP: Image Understanding 60 (1), 26–43.
K ERWIN , W.S., P RINCE , J.L. (1998). Cardiac material markers from tagged MR images. Medical Image
Analysis 2 (4), 339–353.
K RAMER , C., ROGERS , W., T HEOBALD , T., P OWER , T., P ETRUOLO , S., R EICHEK , N. (1996). Remote
noninfarcted regional dysfunction soon after first anterior myocardial infarction: A magnetic resonance
tagging study. Circulation 94, 660–666.
K RAMER , C., et al. (1993). Regional differences in function within noninfarcted myocardium during left
ventricular remodeling. Circulation 88, 1279–1288.
References 589
K YRIAKOU , S., D AVATZIKOS , C. (1998). A biomechanical model of soft tissue deformation with applica-
tions to non-rigid registration of brain image with tumor pathology. In: Medical Image Computing and
Computer Assisted Intervention. In: Lecture Notes in Comput. Sci. 1496 (Springer, Berlin), pp. 531–538.
L ISEIKIN , V.D. (1999). Grid Generation Methods (Springer-Verlag, Berlin).
M AILLOUX , G.E., B LEAU , A., B ERTRAND , M., P ETITCLERC , R. (1987). Computer analysis of heart mo-
tion from two-dimensional echocardiograms. IEEE Trans. Biomed. Engrg. 34 (5), 356–364.
M ARCUS , J., G OTTE , M., ROSSUM , A.V., K UIJER , J., H EETHAAR , R., A XEL , L., V ISSER , C. (1997).
Myocardial function in infarcted and remote regions early after infarction in man: Assessment by magnetic
resonance tagging and strain analysis. Magnetic Resonance in Medicine 38, 803–810.
M C E ACHEN , J., O WEN , R., D UNCAN , J. (1997). Shape-based tracking of left ventricular wall motion. IEEE
Trans. Medical Imaging 16 (3), 270–283.
M C V EIGH , E.R. (1998). Regional myocardial function. Cardiology Clinics 16 (2), 189–206.
M EUNIER , J. (1998). Tissue motion assessment from 3D echographic speckle tracking. Phys. Med. Biol. 43,
1241–1254.
M EYER , F.G., C ONSTABLE , R.T., S INUSAS , A.J., D UNCAN , J.S. (1996). Tracking myocardial deformation
using phase contrast MR velocity fields: A stochastic approach. IEEE Trans. Medical Imaging 15 (4).
M IGA , M., PAULSEN , K., K ENNEDY , F., H OOPES , J., H ARTOV, A., ROBERTS , D. (1998). Initial in-vivo
analysis of 3D heterogeneous brain computations for model-updated image-guided neurosurgery. In: Med-
ical Image Computing and Computer Assisted Intervention, pp. 743–752.
M IGA , M., ROBERTS , D., K ENNEDY , F., P LATENIKI , L., H ARTOV , A., L UNN , K., PAULSEN , K. (2001).
Modeling of retraction and resection for intraoperative updating of images. Neurosurgery 49 (1), 75–85.
N ABAVI , A., B LACK , P., G ERING , D.T., et al. (2001). Serial intraoperative MR imaging of brain shift.
Neurosurgery 48 (4), 787–798.
N AYLER , G., F IRMIN , N., L ONGMORE , D. (1986). Blood flow imaging by cine magnetic resonance.
J. Comp. Assist. Tomog. 10, 715–722.
OMI (1997). Operation of the Mayfield® Accis™ Stereotactic Workstation (OMI® Surgical Products).
O PPENHEIM , A.V., S CHAFER , R.W. (1975). Digital Signal Processing (Prentice Hall, Englewood Cliffs,
NJ).
PAPADEMETRIS , X., O NAT , E.T., S INUSAS , A.J., D IONE , D.P., C ONSTABLE , R.T., D UNCAN , J.S. (2001).
The active elastic model. In: Information Processing in Medical Imaging, IPMI’01, Davis, CA. In: Lecture
Notes in Computer Science 2082 (Springer-Verlag, Berlin).
PAPADEMETRIS , X., R AMBO , J.V., D IONE , D.P., S INUSAS , A.J., D UNCAN , J.S. (1998). Visually interac-
tive cine-3D segmentation of cardiac MR images. J. Am. Coll. of Cardiology 31 (2) (Suppl. A).
PAPADEMETRIS , X., S INUSAS , A.J., D IONE , D.P., C ONSTABLE , R.T., D UNCAN , J.S. (2002). Estimation
of 3D left ventricular deformation from medical images using biomechanical models. IEEE Trans. Medical
Imaging 21 (7), 786–800.
PAPADEMETRIS , X., S INUSAS , A.J., D IONE , D.P., D UNCAN , J.S. (2001). Estimation of 3D left ventricular
deformation from echocardiography. Medical Image Analysis 5 (1), 17–29.
PAPOULIS , A. (1991). Probability, Random Variables and Stochastic Processes, third ed. (McGraw–Hill,
New York).
P ELC , N.J. (1991). Myocardial motion analysis with phase contrast cine MRI. In: Proc. 10th Annual SMRM,
San Francisco, p. 17.
P ELC , N.J., H ERFKENS , R., P ELC , L. (1992). 3D analysis of myocardial motion and deformation with phase
contrast cine MRI. In: Proc. 11th Annual SMRM, Berlin, p. 18.
P ELC , N., H ERFKENS , R., S HIMAKAWA , A., E NZMANN , D. (1991). Phase contrast cine magnetic resonance
imaging. Magn. Res. Quart. 7 (4), 229–254.
P ETERS , T., D AVEY , B., M UNGER , P., C OMEAU , R., E VANS , A., O LIVIER , A. (1996). Three-dimensional
multimodal image-guidance for neurosurgery. IEEE Trans. Medical Imaging 15 (2), 121–128.
P RINCE , J.L., M C V EIGH , E.R. (1992). Motion estimation from tagged MR image sequences. IEEE Trans.
Medical Imaging 11, 238–249.
S ERMESANT , M., C OUDIRE , Y., D ELINGETTE , H., AYACHE , N. (2002). Progress towards an electro-
mechanical model of the heart for cardiac image analysis. In: IEEE Int. Symp. Biomedical Imaging, pp. 10–
14.
S HI , P., S INUSAS , A.J., C ONSTABLE , R.T., R ITMAN , E., D UNCAN , J.S. (2000). Point-tracked quantitative
analysis of left ventricular motion from 3D image sequences. IEEE Trans. Medical Imaging 19 (1), 36–50.
S KRINJAR , O., D UNCAN , J. (1999). Real time 3D brain shift compensation. In: Information Processing in
Medical Imaging, IPMI’99, pp. 42–55.
S KRINJAR , O., N ABAVI , A., D UNCAN , J.S. (2002). Model-driven brain shift compensation. Medical Image
Analysis 6 (4), 361–373.
S KRINJAR , O., TAGARE , H., D UNCAN , J. (2000). Surface growing from stereo images. In: Computer Vision
and Pattern Recognition, CVPR’2000, Hilton Head Island, SC, USA (IEEE Computer Society), pp. 571–
576.
S ONG , S., L EAHY , R. (1991). Computation of 3D velocity fields from 3D cine CT images. IEEE Trans.
Medical Imaging 10, 295–306.
S PENCER , A. (1980). Continuum Mechanics (Longman, London).
S TOKKING , R. (1998). Integrated visualization of functional and anatomical brain images. PhD thesis, Uni-
versity Utrecht.
S TRANG , G. (1986). Introduction to Applied Mathematics (Wellesley–Cambridge Press, Wellesley, MA).
VALLIAPPAN , A. (1981). Continuum Mechanics Fundamentals (Balkema, Rotterdam).
VAN D IJK , P. (1984). Direct cardiac NMR imaging of heart wall and blood flow velocity. J. Comp. Assist.
Tomog. 8, 429–436.
WARFIELD , S.K., TALOS , F., T EI , A., et al. (2002). Real-time registration of volumetric brain MRI by
biomechanical simulation of deformation during image guided neurosurgery. Comput. Visual. Sci. 5.
W EDEEN , V. (1992). Magnetic resonance imaging of myocardial kinematics: Technique to detect, localize
and quantify the strain rates of active human myocardium. Magn. Reson. Med. 27, 52–67.
Y OUNG , A.A., K RAITCHMAN , D.L., D OUGHERTY , L., A XEL , L. (1995). Tracking and finite element analy-
sis of stripe deformation in magnetic resonance tagging. IEEE Trans. Medical Imaging 14 (3), 413–421.
Z HU , Y., D RANGOVA , M., P ELC , N.J. (1997). Estimation of deformation gradient and strain from cine-PC
velocity data. IEEE Trans. Medical Imaging 16 (6).
Methods for Modeling and Predicting
Mechanical Deformations of the Breast
under External Perturbations
Fred S. Azar
Department of Imaging and Visualization, Siemens Corporate Research,
755 College Road East, Princeton, NJ 08540, USA
E-mail: Fred.Azar@scr.siemens.com
Dimitris N. Metaxas
Center for Computational Biomedicine Imaging and Modeling, Division of Computer
and Information Sciences, Rutgers The State University of New Jersey,
110 Frelinghuysen Road, Piscataway, NJ 08854-8019, USA
E-mail: dnm@cs.rutgers.edu
Mitchell D. Schnall
Department of Radiology, Hospital of the University of Pennsylvania,
MRI Bldg. 1 Founders, 3400 Spruce St., Philadelphia, PA 19104, USA
E-mail: schnall@oasis.rad.upenn.edu
List of symbols
M Mass matrix
D Damping matrix
K Model stiffness matrix
Ke Elemental stiffness matrix
q Vector which contains the displacement degrees of freedom
qi = (qi,x , qi,y , qi,z ) Vector-displacement degrees of freedom for node i in the model
591
592 F.S. Azar et al.
Xi = (xi , yi , zi ) Vector-position of node i in the model

qe Element nodal displacement vector
gq Inertial forces vector
fq Generalized external forces vector
fi,internal Internal stiffness force vector on node i
fe Element force vector
µ Mass density
Ve Volume of element
σ Nodal stress vector
ε Nodal strain vector
εn Strain function for tissue type n
σn Stress function for tissue type n
u = u(x, y, z), v = v(x, y, z), w = w(x, y, z) Displacement fields in the x, y and z
directions, respectively
En Elastic modulus, function of strain for tissue type n
E Lagrangian strain tensor
F Deformation gradient tensor
1. Introduction
Breast cancer is the second leading cause of cancer deaths in women today (after lung
cancer) and is the most common cancer among women, excluding non-melanoma skin
cancers. According to the World Health Organization, more than 1.2 million people
were diagnosed with breast cancer in 2001 worldwide. The American Cancer Society
estimated approximately 192,000 new diagnosed cases of invasive breast cancer (stages
I–IV), and 41,000 deaths from breast cancer among women in the United States in
2001. The incidence rate of breast cancer (number of new breast cancers per 100,000
women) increased by approximately 4% during the 1980s but leveled off to 100.6 cases
per 100,000 women in the 1990s (F ERLAY, B RAY, et al. [2001]). The death rates from
breast cancer also declined significantly between 1992 and 1996, with the largest de-
creases among younger women. Medical experts attribute the decline in breast cancer
deaths to earlier detection and more effective treatments. While breast cancer is less
common at a young age (i.e., in their thirties), younger women tend to have more ag-
gressive breast cancers than older women, which may explain why survival rates are
lower among younger women. The standard treatment against breast cancer today is
to cut out either the tumor or the whole affected area. The only way today to find out
for sure if a breast lump or abnormal tissue is cancer, is by having a biopsy. The sus-
picious tissue, which is removed by a surgeon or radiologist during a biopsy, is then
examined under a microscope by a pathologist who makes the diagnosis. A biopsy is
done most of the time (except for palpable lesions) with the help of images of the breast
obtained using an imaging technique, such as X-rays (PARKER , L OVIN , J OBE , et al.
[1990], D RONKERS [1992], D ERSHAW [1996]) or Magnetic Resonance Imaging (MRI)
(F ISCHER , VOSSHENRICH , K EATING , B RUHN , D OLER , O ESTMANN and G RABBE
[1994], O REL , S CHNALL , N EWMAN , P OWELL , T OROSIAN and ROSATO [1994], F IS -
CHER , VOSSHENRICH , D OLER , H AMADEH , O ESTMANN and G RABBE [1995]).
Methods for modeling and predicting mechanical deformations 593
Magnetic resonance imaging uses radio waves and magnetic fields to diagnose dis-
eases. Patients are asked to lie on a table during the test, which takes about 30 minutes.
They are then advanced into the MRI machine, which contains a strong magnetic field
(1.5–4 T). The method consists of injecting a contrast-enhancing dye-like material into
the patient’s bloodstream and using magnetic resonance imaging to monitor the way in
which this material is taken up and cleared out by the tumor tissue. The ability to identify
a mass in the breast requires the mass to have a different appearance (or a different con-
trast) from normal tissue. With MRI, the contrast between soft tissues in the breast is 10
to 100 times greater than that obtained with X-rays (F ISCHER , VOSSHENRICH , K EAT-
ING , B RUHN , D OLER , O ESTMANN and G RABBE [1994], H ARMS and F LAMIG [1994],
O REL , S CHNALL , N EWMAN , P OWELL , T OROSIAN and ROSATO [1994], F ISCHER ,
VOSSHENRICH , D OLER , H AMADEH , O ESTMANN and G RABBE [1995]). There are
additionally multiple reports of MR imaging-detected breast cancers that are mammo-
graphically, ultrasonographically, and clinically occult (H ARMS , F LAMIG , et al. [1993],
B OETES , BARENTSZ , et al. [1994], G ILLES , G UINEBRETIERE , et al. [1994]). Fig. 1.1
shows an MR image of a breast revealing areas of cancer spread (arrows) in addition to
the larger tumor. The additional areas of cancer were not visible on the patient’s X-ray
mammograms.
The main disadvantage of breast MRI is its cost, which today is about 5 times that
of X-ray mammography. According to most physicians, mammography is used more
widely for breast cancer detection because it is a inexpensive technique, and most health
institutions can afford to buy mammographic equipment. Today, however, an area of
technical development is in the field of low-cost, dedicated breast MR systems, which
could reduce the cost of breast MR imaging dramatically. Sales of magnetic resonance
imaging (MRI) scanners reached an all-time high in 1999, topping the $1 billion mark.
MRI Industry Report, a quarterly newsletter published by Miller Freeman, reported that
F IG . 1.1. MRI of a breast showing mammographically occult cancer tumors.

928 new MRI systems, valued at $1.07 billion, were installed at US hospitals and clin-
ics in 1999. Sales revenue rose 12.2% due to the strong demand for new MRI systems
introduced in 1998 and 1999. Sales appear to be on the rise for more costly MRI sys-
tems that feature more and improved capabilities, while sales for less costly systems are
tapering off.
In MR breast imaging, the specificity has been reported to vary between 37 and 97%
(H ARMS , F LAMIG , et al. [1993], B OETES , BARENTSZ , et al. [1994], G ILLES , G UINE -
BRETIERE , et al. [1994]): not all contrast-enhancing lesions prove to be malignant. With
such a technique, which is highly sensitive, but not highly specific, an image-guided
breast localization and biopsy system is needed to help differentiate between the benign
enhancing lesions and carcinomas (O REL , S CHNALL , N EWMAN , P OWELL , T OROSIAN
and ROSATO [1994]). A whole-body MR system at 1.5 T (Signa; GE Medical Systems,
Milwaukee, WI) is used for all needle localizations. The technique requires that the
patient lies prone with the breast gently compressed between medial and lateral plates
(Fig. 1.2). A multicoil array is used, with two coils placed on the medial plate and one
coil on the lateral plate. The lateral plate contains a grid of approximately 1000 18-
gauge holes placed at 5 mm intervals, which guide the needle in a plane parallel to the
tabletop. The correct hole in the plate is identified and a needle is inserted through that
hole into the breast with a pre-calculated depth (Fig. 1.2).
The MR imaging-guided localization technique encounters the following problems:
• The appearance, size and shape of the potential cancer lesion greatly depend on
the dynamics of the interaction between the patient’s physiology and the contrast-
enhancing agent (T OFTS , B RIX , et al. [1999]). The lesion may clearly appear only
in the two minutes following the contrast agent injection, then the signal intensity
may vary arbitrarily, and it is possible that the apparent boundaries of the lesion
may change dramatically.
• The needle is not a very sharp object and cannot be smoothly inserted in the
breast. Every time the tip of the needle reaches the interface between two dif-
ferent types of tissue, its further insertion will push the tissue instead of pierc-
ing it, causing unwanted deformations until the pressure on the tissue interface
is high enough. As soon as the needle pierces the displaced tissue interface,
that interface quickly settles back to its original position, or somewhere close
to it. The best way to remedy this problem would be to compress the breast as
much as possible, which would minimize internal deformations. However, do-
ing that would cause blood to be squeezed out of the breast, and would dra-
matically alter the appearance and shape of the lesion on the MR image (the
perfusion changes to the lesion would disrupt Gd uptake), without mentioning
the high level of discomfort for the patient who would be very reluctant to feel
the pain for the entire duration of the procedure. The best solution would be
to mildly compress the breast and obtain MR images clearly showing the po-
sition, shape and extent of the lesion. Then the breast would be highly com-
pressed in order to minimize internal deformations during the needle insertion
only. However, the missing link is to predict the displacement of the lesion
from the mildly compressed configuration, to the highly compressed configura-
tion.
F IG . 1.2. Breast compression setup before a MR-guided needle localization procedure (top). The patient lies
prone on the bed with her breast gently compressed between medial and lateral plates, as the bed is being
moved into the high-field magnet (bottom).
The above limitations coupled with the deformable structure of the breast makes nee-
dle procedures very sensitive to the initial placement of the needle and to the amount
of breast compression. It thus becomes relatively uncertain that the tissue specimen re-
moved during the biopsy procedure actually belongs to the lesion of interest, due to
the added difficulty of accurately locating the tumor’s boundaries inside the breast. It
is therefore important to develop techniques, which would solve or bypass the afore-
mentioned problems, increase the level of confidence of a biopsy result (improving the
diagnosis), and decrease the cost to society (including health care expenses). The finan-
cial savings could be considerable, and the morbidity associated with the biopsy proce-
dure including the lost time from work that occurs as a result of biopsy could be greatly
reduced (S TAVROS , T HICKMAN , R APP, D ENNIS , PARKER and S ISNEY [1995]).
We present a virtual deformable breast model of the patient whose geometry is con-
structed from MR data. The elastic properties of the deformable model are based on
the use of finite elements with nonlinear material properties capable of modeling the
deformation of the breast under external perturbations. A high-field 1.5 Tesla machine
Signa Horizon Echospeed (GEMS, Milwaukee) is used to obtain the 3D breast image
sets. The image sets are used to construct the geometry of the finite element model.
Contours of the breast are extracted, and each breast slice is segmented to locate the
different tissue types, using appropriate custom-written software. The model geometry
is then created using a custom-written preprocessor, which allows for a variable mesh
size. We also developed a software algorithm (BreastView), which models large defor-
mations of the breast model depending on the desired accuracy of the deformation. We
hypothesize that the structural complexity of the breast can be simplified to only assign
to the model elements, an average value of the mechanical properties of glandular, fatty,
and cancerous tissue.
The major novelties in this model include the following:
• Breast plate compression results in a large compression, meaning that the total
distance between the two plates decreases by more than 10%. In order to model
such large deformations, we divide the large deformation compression into a num-
ber of much smaller displacement steps. For every displacement step, we make
use of small strain theory. Strain is calculated using Cauchy’s infinitesimal strain
tensor formula (F UNG [1994]). After every small displacement iteration, the tis-
sues’ different material properties are recalculated in all model elements whose
maximum principal strain has changed, in order to model the materials’ non-
linear behavior. The main advantage of using small strain formulation lies in its
simplicity, ease of implementation and fast computation. However, being an in-
cremental formulation, it could lead to an accumulation of discretization errors
and in consequence, to a lack of accuracy (S ZEKELY, B RECHBUHLER , H UTTER ,
R HOMBERG and S CHMID [1998]). A way to solve this potential problem would
be to use a total Lagrange formulation (as the one we use for estimating the non-
linear material properties of tissue from one displacement iteration to the next),
in which every state is related to the initial configuration. However, that would
complicate the formulation and slow down the computation of the solution. We
show in a silicon phantom study that the incremental errors introduced by small
strain formulation can be neglected for the purpose of this model and overall
study.
• We present a new breast fatty tissue material model, which takes into account the
effect of fat compartmentalization due to Cooper’s ligaments in the breast. We
show through empirical evidence that fat compartmentalization occurs as the breast
is being compressed, and that the new updated fatty tissue material model takes
that phenomenon into account, and performs better than the original fatty tissue
model.
• We apply finite element modeling theory to model the deformation of a human
female breast in such a way that the entire process takes less than a half-hour
(compared to several hours using a commercial finite element modeling pack-
age), which according to the clinicians consulted, is a reasonably short time du-
ration.
This model can be used effectively in several different applications:
• A new method for guiding clinical breast biopsy. This method involves imaging the
patient’s breast without any or little compression before a needle procedure, then
compressing the breast, and its virtual finite element model (by applying the same
pressure to both), and using the displacement of the virtual tumor model to predict
the displacement of the real cancer tumor. It is important to note that during the
entire procedure (imaging, needle localization, and/or biopsy), the patient remains
in the same prone position, and only the equipment ‘moves’ around the patient.
Therefore, perturbations caused by the patient’s movements are minor. A model
like the one presented here is important to this procedure, in which any improve-
ment in confidence for localizing the cancer tumor could become life-saving.
• Other applications. A finite element model of the breast can be a very flexible tool
for many applications including registration of different breast MR data sets of the
same patient taken under different compression amounts, or registration of differ-
ent data sets from different imaging modalities. Other possible applications include
diagnosis, measurements, surgery planning, simulations of deformation due to in-
serting a needle, and further away, virtual surgery and tele-surgery. It is important
to note that the current model is based on a prone-acquired MRI and virtually all
breast surgery is performed in the supine position. Therefore, the rotational invari-
ance of the model would need to be proven for prone-supine deformations.
2. Organization of the chapter
The chapter is organized as follows:

1. In Section 3, the related work is presented and compared to our model. The general
flowchart of operation is described.
2. The following six sections present the methods used, going through image ac-
quisition and data extraction (Section 4), 3D mesh domain creation (Section 5),
model dynamics (Section 6), internal model forces (Section 7), large deformation
models (Section 8), and non-linear material models (Section 9).
3. The following four sections describe phantom and patient studies in detail:
(a) A phantom study (Section 10). A deformable silicon gel phantom was built
to study the movement of a stiff inclusion inside a deformable environment
under plate compression. The phantom was imaged undeformed, then com-
pressed (14%). The performance of our software algorithm was compared to
that of a robust commercial FEM software package. A 3D deformable model
of the phantom was built from the resulting MR data using our custom-
written software and was virtually compressed using BreastView. Another
FEM was built using a commercial pre-processor (PATRAN, MSC, CA) from
the phantom’s directly measured dimensions, and was virtually compressed
using ABAQUS (HKS, Rhode Island). The displacement vectors of the 8
corners of the stiff inclusion and its center were measured both from the MR
images and from the two finite element models.
(b) A patient study (Section 11). A patient’s breast was imaged uncompressed
and then compressed 26%. The corresponding deformable model was built
and was virtually compressed to match the real compression amount. We
tracked the displacement of a small cyst inside the patient’s breast, and used
the deformable model to predict the cyst’s position in the real compressed
breast. We also tracked the displacement of two vitamin E pills taped to the
surface of the patient’s breast. We present a convergence analysis and a ma-
terial properties sensitivity analysis. The results show that it is possible to
create a deformable model of the breast based on the use of finite elements
with non-linear material properties, capable of modeling the deformation of
the breast in a clinically useful amount of time (less than a half-hour for the
entire procedure).
(c) A clinical breast compression study (Section 12). Three patient breasts were
imaged uncompressed and then compressed. Patients were chosen to have a
variety of breast cancers of different shape, size and location in the breast.
The displacement of the cancer tumors was recorded in the three patient
breasts. A model of each patient’s breast was constructed, and then used to
predict the deformation and displacement of the cancer tumors after breast
compression.
(d) Registration of breast MR images of the same patient under different com-
pressions (Section 13). A patient’s breast was imaged under two different
plate compression amounts. A deformable model of that breast was con-
structed and compressed using virtual compression plates. The displacement
of a small cyst was recorded in the real breast and compared to the displace-
ment of the ‘virtual cyst’ in the deformable model from one compression
state to the other.
4. The following two sections deal with additional issues, including potential sources
of error (Section 14) and specific properties of reliability (Section 15).
5. Section 16 summarizes the major novelties in the model, and Section 17 presents
the concluding remarks.
6. Finally, Section 18 (Appendix) presents the finite element modeling theory in de-
tail, as well as details of the silicon phantom construction.
3. Related work
Finite element modeling has been used in a very large number of fields. However it
is only recently that deformable models have been used to simulate deformations in
soft tissue. Physical models are among the first to be used. Among these physical mod-
els, elastic (linear and visco-elastic) models have been extensively described in the lit-
erature (C HEN and Z ELTZER [1992], S PEETER [1992], R EDDY and S ONG [1995]).
The most widely used representations for deformable volumes are parametric models
with B-spline representation (Z IENKIEWICZ [1977]). Other possible models are mass–
spring models (M ILLER [1988], C HADWICK , H AUMANN and PARENT [1989], L U -
CIANI , J IMENEZ , F LORENS , C ADOZ and R AOULT [1991], N ORTON , T URK , BACON ,
G ERTH and S WEENEY [1991], J OUKHADAR [1995]) and implicit surfaces (D ESBRUN
and G ASCUEL [1995]). The mass–spring methods have been used most of the time
for surgery simulation due to their simplicity of implementation and their lower com-
putational complexity (K UEHNAPFEL and N EISIUS [1993], BAUMANN and G LAUSER
[1996], M ESEURE and C HAILLOU [1997]). Other methods have relied on geometry
rather than physics to predict breast deformation (B EHRENBRUCH , M ARIAS , A R -
MITAGE , YAM , M OORE , E NGLISH and B RADY [2000]).
Finite element models are less widely used due to the difficulty of their implementa-
tion and their larger computing time. There are many powerful commercial FEM pack-
ages that allow complex simulations of deformation such as ABAQUS (HKS, Rhode
Island); breast tissue is relatively complex, and consists of layers of different tissues
interlaced with ligaments and fascias. Very complex models would be needed to model
these objects realistically. However, the complexity of the model and the required com-
putational time (which can extend to several days on a SGI workstation) would prohibit
these models from being useful clinically. Few models of the breast have actually been
implemented using the commercial software packages, and have mostly involved phan-
tom studies (A ZAR , M ETAXAS and S CHNALL [1999], S CIARETTA , B ISHOP, S AMANI
and P LEWES [1999], W ILLIAMS , C LYMER and S CHMALBROCK [1999]). A real-time
system has been recently developed for hepatic surgery simulation, and involves de-
formations of soft tissue (C OTIN , D ELINGUETTE and AYACHE [1999]). An explicit
large displacement model for interactive surgery simulation using fast non-linear finite
element models was proposed by P ICINBONO , D ELINGUETTE and AYACHE [2001],
and models for physically realistic simulation of global deformation were proposed by
Z HUANG and C ANNY [1999]. Finite element models were also proposed for bone and
muscle biomechanics research (H OU , L ANG , et al. [1998], K ABEL , R IETBERGEN , et
al. [1999] Y UCESOY, KOOPMAN , et al. [2002]).
The desired accuracy of the deformation in the breast model must therefore be bal-
anced against the need for speed. In order to develop a model for deformation of the
breast, we must decide on a geometric description of the breast, a mathematical model
of the elastic deformation, and a solution algorithm that is both fast and yields a reason-
ably realistic result. For that reason, what matters most is that the breast deformation
results be realistic and available in a clinically useful time (less than thirty minutes
for the entire procedure), and that the model be robust and show a consistent and pre-
dictable behavior. In our approach, we strive to integrate the requirements for a realistic
simulation of deformation, and the reasonably fast time modeling which is a necessity
if the model is to be used in a clinical environment: by incorporating the geometric de-
finition of the breast model into the physics-based framework developed in M ETAXAS
[1992], M ETAXAS and T ERZOPOULOS [1993], we create a deformable breast model,
capable of reasonably predicting the internal deformations of a real patient breast after
plate compression (A ZAR , M ETAXAS and S CHNALL [2001], A ZAR , M ETAXAS and
S CHNALL [2002]).
Fig. 3.1 shows the general flowchart of how the FEM of the breast is created (vertical
dimension), and the process by which it is virtually compressed (horizontal dimension).
We start with the patient’s breast MR data which constitutes a 3D image set of parallel
slices. Each of these slices is segmented and classified into different tissue types (glan-
dular, fatty and/or cancerous) semi-automatically. Given this data, the 3D mesh of the
model is generated automatically, and every element in the mesh is assigned a particular
tissue type value. Then the deformation process starts from the given initial and bound-
ary conditions. The large compression plate displacement is translated into a number
of much smaller displacement iterations, each of which is applied in turn to the FEM
of the breast. For each applied displacement iteration, the equation of motion must be
solved and the intermediate model node displacements are calculated. Following that
step, the elasticity value of every element in the model is updated given its principal
strain value and the non-linear material model of the tissue type it represents. After all
the displacement iterations are applied the final compressed model of the breast is ob-
F IG . 3.1. General flowchart of operation.

tained. Because of the high variability of breast shapes and the deformation of the breast
when compressed in order to be imaged, the model devised closely follows the contours
of the patient breast.
4. Image acquisition and data extraction
The patient data is a set of parallel 2D spoiled gradient echo MR axial slices of the
breast. Usually, an axial T1-weighted spin echo sequence is performed with a repetition
time of 500 msec, and an echo time of 12 msec, with a 12–16 cm field of view, a 1–3 mm
thick section, and a 256 × 256 matrix. The acquisition ensures a 3D visualization of the
patient’s breast. First, the MR image 3D set is converted into a set of axial slices (if
the original data is not axial) through automatic resampling of the data using a software
package such as Scion Image (Scion Corp., Maryland).
The MR images are loaded to a program, which enables segmentation through con-
tours. The contours of the breast boundary are created semi-automatically (using Scion
Image and Adobe Photoshop) using a threshold-based technique, and saved to a file
for subsequent use. At the same time each MR image is automatically segmented into
parenchyma, fat and/or lesion tissue, using a robust segmentation algorithm (each tissue
type is assigned a specific gray level value).
The segmentation algorithm used is based on the concept of fuzzy connected-
ness (C ARVALHO , G AU , H ERMAN and KONG [1999], S AHA , U DUPA and O DHNER
[2000]). This approach is semiautomatic, in the sense that when using the segmentation
program (developed by Carvalho et al.) the user must identify seed voxels, which defi-
nitely belong to the various objects in the image (fat, glandular, cancerous tissues). The
user-selected seed voxels are then used for automatic segmentation of the entire image.
This algorithm yields accurate and robust results, and is able to segment a breast full
MR 3D volume in less than a minute (on a standard PC workstation), after only spend-
ing a few minutes selecting the seed voxels. Although the technique works reasonably
well due to the detection of different tissue contrasts, it may not segment lesions well
because it is not parametric enough with respect to the pharmacokinetics of the Gadolin-
ium contrast-enhancing agent.
5. 3D mesh domain creation
BreastView, a custom-written program in C, takes as input the set of breast contours

and the segmentation results. Running on a SGI workstation, it generates the 3D com-
putational domain (mesh) of the breast in a few seconds, allowing to scale the volume
elements to any size. The program can also generate a finite element model (FEM) file
readable by FEM software ABAQUS. This file contains the definition of the volume
elements, as well as the boundary conditions, and the different material properties. Two
types of 3D volume elements are used in the model (Fig. 5.1, bottom). Solid 8-node
(hexahedral) trilinear isoparametric elements are used to model all breast tissues except
for the skin, and are assigned non-linear material properties. 3-node triangular isopara-
metric elements are used to model the skin, and are assigned non-linear elastic proper-
ties in the plane of their triangular surface, and no axial stiffness. The nodes forming the
F IG . 5.1. Slice contour with 2D-mesh generation (top), and elements used to build the deformable model
(8-node element, 3-node triangle, bottom).
2D mesh representing the skin are the same nodes belonging to the hexahedral elements
at the borders of the breast model.
BreastView allows the mesh to be as dense as needed. These volume elements are
well-suited for the purpose of the model: if a physician only needs approximate infor-
mation on how the tumor is moving, the tumor could be included in one volume element
(by correctly scaling the mesh density) which would be assigned its own mechanical
characteristics. Fig. 5.1 (top) shows a contour and the 2D mesh generated in it.
In order to create the 3D mesh, we first find the principal direction d for the entire
contour set (direction of the v-axis, Fig. 5.1, top), from the chest wall to the nipple. Any
line with direction d in the plane of a contour, which intersects the contour, will intersect
it twice. This direction is that of the line orthogonal to the line, which passes through
the two endpoints (the principal direction method using the eigenvectors of the contour
set of points did not yield the best direction d). The principal direction d for the entire
contour set is obtained by averaging the computed individual slice directions. The center
of gravity G of the contour set is calculated. The 3D mesh can now be easily generated
following the U (along u) and V (along v) resolution desired, in the orthogonal (u, v)
basis centered at G. This algorithm ensures that we have the same number of points on
every V-line, and the same number of V-lines on every slice.
In order to avoid that degenerate elements at the edge of the breast may be created
during the 3D meshing process, we introduce a predetermined offset for every slice,
which is a percentage of the length L of the breast slice, as shown in Fig. 5.1. The row
of element nodes generated closest to the edge of the breast slice is at a distance of
L × Offset from the edge of the breast. Finally, every node on this final row is projected
along the v-axis onto the edge of the breast slice.
We implemented the meshing in the z-direction (number of slices) in two different
ways:
• There can be as many element slices (in the z-direction) as there are segmented
data slices.
• We make the meshing in the z-direction variable, increase the mesh density around
the point of interest, and decrease it away from the point of interest. This scheme
can enable us to decrease the total number of model slices, and hence greatly ac-
celerate the simulation.
Since the point of interest is usually a hard inclusion such as a cyst or a tumor, it
is more important for the mesh to be denser in that area than in the slices away from
the inclusion, which may have less impact on the displacement of the inclusion. By
implementing such an algorithm, we can also decrease the total number of model slices,
thereby greatly decreasing the compression simulation time.
Once the 3D mesh is created, each element is assigned a material property corre-
sponding to that of fat, parenchyma or cancerous tissue. In order to determine what
type of material property to assign to any one 3D element, an algorithm makes use of
the automatically segmented breast MR data. The algorithm calculates what percent-
age of fat, parenchyma and cancerous tissue lie within the element, by identifying the
corresponding number of voxels which lie in the element. The material property that
corresponds to the highest percentage is the one assigned to the element. It is clear from
this algorithm, that the finer the mesh, and the more precise the assignment of material
properties will be.
6. Model dynamics
The governing Lagrange equations of motion are second order differential equations
given as follows (T ERZOPOULOS , P LATT, et al. [1987], M ETAXAS [1992], M ETAXAS
and T ERZOPOULOS [1993]):
∂ 2q ∂q
M +D + Kq = gq + fq , (6.1)
∂t 2 ∂t
where M, D and K are the mass, damping and stiffness matrices, respectively. The
vector q contains the displacement degrees of freedom, gq are the inertial forces arising
from the dynamic coupling between the local and global degrees of freedom, and fq are
the generalized external forces associated with the model’s degrees of freedom.
In our case, given the nature of the problem, and the fact that we consider sta-
tic deformations only, it makes sense to simplify the equations of motion by setting
the mass density to zero, which will still preserve useful first-order dynamics that
achieve equilibrium. Setting the mass density µ to zero causes the mass matrix M
and the inertial forces gq to vanish (PARK , M ETAXAS , YOUNG and A XEL [1996a],
H ABER , M ETAXAS and A XEL [2000]). This results in the first-order dynamic system
(PARK , M ETAXAS , YOUNG and A XEL [1996b]),

∂q
D + Kq = fq . (6.2)
∂t
Because these equations lack an inertial term, the system comes to rest as soon as all
the forces equilibrate or vanish. We use D as a stabilizing factor only, and do not impose
physical damping, which cannot be measured from experimental data. Therefore, we
assume that D is diagonal and constant over time, and set it equal to the identity matrix
(D = I). Therefore, the breast model is deformed using the following reduced form of
Lagrange’s equation of motion:
∂q
αI + Kq = fq , (6.3)
∂t
where the vector q contains the displacement degrees of freedom, α is a numerical
damping constant and the vector fq contains the total external forces due to body forces
which in our case would be gravity. We are modeling the compression of the breast for
a patient in a prone position. The breast is immobile in a state of equilibrium (the skin
is mainly responsible for counteracting the effects of the gravitational force), and the
plates compressing the breast move in a direction orthogonal to that of the gravitational
forces. Therefore the vector fq can be set to zero.
We may approximate the equation by using the finite element method described pre-
viously. Through this method, all quantities necessary for the Lagrange equations of
motion are derived from the same quantities computed independently within each finite
element. The various matrices and vectors involved in the Lagrange equations of motion
are assembled from matrices computed within each of the elements.
Therefore, in our algorithm we do not assemble the model stiffness matrix K, but
work with the elemental stiffness matrices Ke , and assemble the forces around the
nodes. Compared to the classical theory, this method is equivalent to a relaxation
method, i.e., calculating the residuals at the nodes.
Stiffness (represented by the K matrix) can also be viewed as an internal resistance
which results in a force equal to −Kq. Lagrange’s equation of motion can then be writ-
ten for each and every node i in the model as (H ABER , M ETAXAS and A XEL [2000])
∂qi
αI = fi,total = fi,internal, (6.4)
∂t
where qi is the 3D nodal displacement at node i, fi,internal = [−Kq]i is an internal
stiffness force. The nodal equation becomes
∂qi
αI = fi,internal, (6.5)
∂t
where α is a numerical damping constant, which we set to 1. The time-discretized nodal
equation can then be written as
qi,t +1 = qi,t + t · fi,internal . (6.6)
At each displacement iteration, we integrate the equation using an Euler method with
adaptive step sizing. The step size t is inversely proportional to the integration error.
Since the forces multiply the time step, the adaptive step sizing effectively modulates
the strength of the forces, thereby making the solution converge much faster than with
a simple Euler integration technique (H ABER , M ETAXAS and A XEL [2000]). We com-
pute two estimates of qi,t +1 , qi,a by taking an Euler step of size t, and qi,b by taking
two Euler steps of size t/2. Since qi,a and qi,b differ from each other by O(t 2 ),
a measure of the current error is e = ni=1 |qi,a − qi,b |, where the error is over the sum
of the displacement errors of all nodes. If we are willing
√ to have an error of as much as
eA , then the new step size can be written as tnew = eA /e · t. The numerical inte-
gration ends when the difference between the sum of all displacements from one time
iteration to the next is less than a predetermined threshold.
7. Internal forces due to stiffness
The degrees of freedom of the model are the 3D displacements of the finite ele-
ment nodes. In our formulation, stiffness is represented as an internal nodal force
fi,internal . The element force vector fe contains the forces on the element nodes: fe =
[fi,internal fi+1,internal . . . fi+n−1,internal]T where fi,internal is the internal force on node i,
and n is the number of nodes in the element. fe is calculated for each element as
fe = Kqe , (7.1)
where qe contains the element’s nodal displacements: qe = [qi qi+1 . . . qi+n−1 ]T . The
elemental stiffness matrix, Ke , incorporates the geometry, material properties of the
element and is a triple integral over the volume Ve of the element. It is computed from

Ke = BT DB dV . (7.2)
Ve
The matrix D contains the material property information (which includes the value
of the material’s elastic modulus Ek (εk )) and relates nodal stresses σ to nodal strains ε:
σ = Dε.
The matrix B relates nodal strains to displacements and incorporates Cauchy’s infin-
itesimal strain tensor formulation (see Eq. (7.1)):
ε = Bqe . (7.3)
Once the element forces are calculated, each element contributes to the total internal
nodal force fi,internal for each of its nodes. A detailed derivation of the finite element
formulation for calculating the stiffness matrix, K, is presented in Z IENKIEWICZ and
TAYLOR [1989].
8. Modeling large deformations
Compressing a breast using compression plates results in a large compression of the

breast, meaning that the total compression distance between the two plates decreases by
more than 10%. In order to model such large deformations, we divide the large defor-
mation compression into a number of much smaller displacement steps. For every dis-
placement step, we make use of small strain theory. Strain is calculated using Cauchy’s
infinitesimal strain tensor formula (given in the unabridged notation (F UNG [1994])):
∂u ∂v ∂w
εxx = , εyy = , εzz = ,
∂x ∂y ∂z

1 ∂u ∂v 1 ∂u ∂w
εxy,yx = + , εxz,zx = + ,
2 ∂y ∂x 2 ∂z ∂x

1 ∂v ∂w
εyz,zy = + . (8.1)
2 ∂z ∂y
u = u(x, y, z), v = v(x, y, z), w = w(x, y, z) are the displacement fields in the x, y
and z directions, respectively, from one small displacement iteration to another.
The main advantage of using small strain formulation lies in its simplicity, ease of
implementation and fast computation. However, being an incremental formulation, it
could lead to an accumulation of discretization errors and in consequence, to a lack
of accuracy (S ZEKELY, B RECHBUHLER , H UTTER , R HOMBERG and S CHMID [1998]).
A way to solve this potential problem would be to use a total Lagrange formulation
(as the one we use for estimating the non-linear material properties of tissue from one
displacement iteration to the next), in which every state is related to the initial configu-
ration. However that would complicate the formulation and slow down the computation
of the solution. We show in the phantom study that the incremental errors introduced
by small strain formulation can be neglected for the purpose of this model and overall
study.
9. Modeling non-linear material properties
Most biological tissues display both a viscous (velocity dependent) and elastic response,
however since we are only interested in slow displacements, the great majority of the
forces developed can be attributed purely to the elastic response. All tissues involved in
the breast can be considered:
• isotropic (H AYES , K EER , H ERMANN and M OCKROS [1972], K ROUSKOP, W HEE -
LER , K ALLEL , G ARRA and H ALL [1998]),
• homogeneous (S ARVAZYAN , S KOVORODA , E MELIANOV, F OWLKES , P IPI , A D -
LER , B UXTON and C ARSON [1995], S KOVORODA , K LISHKO , G USAKYAN ,
M AYEVSKII , Y ERMILOVA , O RANSKAYA and S ARVAZYAN [1995]),
• incompressible (F UNG [1993]),
• to have non-linear elastic properties (F UNG [1981], Z HANG , Z HENG and M AK
[1997]).
With these assumptions, it is possible to define the mechanical behavior of breast
tissue using a single elastic modulus En , which is a function of strain εn for tissue type
n (σn is the stress):
∂σn
En = = f (εn ). (9.1)
∂εn
This non-linear relationship is calculated for every tissue type from uniaxial stress–
strain experiments using tissue samples. The experimental curves are fit to a material
model, which can be characterized using a small number of parameters. In order to
model the non-linear mechanical behavior in every element, the following steps are
followed, after every small deformation increment:
(1) We calculate the Lagrangian strain tensor, E, at the center of the element. E is
a measure of the deformation of a point in the model with respect to its initial
position (a detailed description can be found in S PENCER [1980]):
1 T
E= F F−I . (9.2)
2
F is the (3 × 3)-deformation gradient tensor, and its components can be written
as
∂xp
Fpq = , (9.3)
∂Xq
where xp is one of the three components of x, the final position vector, and Xq is
one of the three components of X, the initial position vector. In our case, the final
position of an element center point is after a given deformation increment, and
the initial center point is before the first deformation increment. I is the identity
matrix.
(2) We calculate the maximum principal strain component, emax , which is given by
the largest eigenvalue of the strain tensor E.
(3) Using the experimental stress–strain curve for the particular tissue type, the el-
ement is assigned a Young’s modulus value corresponding to the slope of the
curve at the strain measure emax : En = f (emax ).
Therefore, after every deformation increment, the element’s stiffness value is updated
to model the material’s non-linear behavior. By doing so, we model a continuously
differentiable stress–strain curve (in the limits of our deformations) as a continuous
step-wise linear curve (Fig. 9.1).
F IG . 9.1. Modeling a non-linear stress–strain curve (left) with a piecewise linear curve (right). Di is the
interval used during the ith deformation increment.
10. A phantom study
10.1. Experimental design and methods
A deformable silicon gel phantom was built to study the movement of a stiff inclusion
inside a deformable environment (as a tumor inside the breast) under plate compres-
sion (A ZAR , M ETAXAS , M ILLER and S CHNALL [2000]). The phantom was imaged
undeformed, then compressed. A 3D deformable model of the phantom was built from
the resulting MR data, and was virtually compressed using custom-written software
(BreastView). Another FEM of the phantom was built using a commercial pre-processor
(PATRAN) from the phantom’s directly measured dimensions, and was compressed vir-
tually using a powerful commercial FEM software package (ABAQUS) which uses im-
plicit integration schemes.
The displacement vectors of the corners of the stiff inclusion and its center were
measured both from the MR images and from the two finite element models. This study
also serves to validate our mathematical model of large deformations by comparing our
results to that of a commercial FEM software package.
The explicit Finite Element formulation, which we implemented, using small strain
theory, may not always be as accurate as the implicit one, and time discretization er-
rors can accumulate. The main reason is that our strain formulation is of an incremental
form; this could lead to a lack of accuracy (KOJIC and BATHE [1987]). The results of
this phantom study show that the compressed model allows us to track the position and
motion of the stiff inclusion in the real compressed deformable environment. Further-
more, after comparing the compressed BreastView and ABAQUS models, the results
also show that using a small strain approximation in our finite element formulation does
not introduce a significant error.
10.1.1. Phantom construction

The phantom was designed to have magnetic properties (T1 and T2) similar to those
of human breast tissue, to withstand large deformations (20% or greater), and to en-
able controlled deformations. The gel phantom was build using Sylgard Dielectric Gel
527 (Dow Corning, Midland, MI). A similar silicon gel (model Q7-2218, Dow Corn-
ing) has been suggested for use in MR imaging (G OLDSTEIN , K UNDEL , DAUBE -
W HITERSPOON , T HIBAULT and G OLDSTEIN [1987]), and the same silicon gel was
used to validate tagging with MR imaging to estimate material deformation (YOUNG ,
A XEL , D OUGHERTY, B OGEN and PARENTEAU [1993]). The geometry of the de-
formable phantom consists of a rectangular box (84 × 82 × 70 mm) containing a rec-
tangular inclusion (20 × 23 × 20 mm), which is 4.3 times stiffer than the surrounding
silicon (Fig. 10.1). For more details about the silicon phantom construction, see Appen-
dix.
10.1.2. MR imaging of silicon phantom

The full silicon gel phantom was placed in a custom-built pressure device, where a pres-
sure plate could compress the gel phantom in a similar way as with a real breast with the
desired amount of deformation. The whole setup was secured firmly and imaged with a
F IG . 10.1. Top view of the setup for imaging the compressed gel phantom (top), and construction of phantom
model: 2D axial slice shown (bottom).
whole body 1.5 T superconducting magnet (GEMS, Milwaukee). The silicon gel phan-
tom was first imaged undeformed. The compression plate then applied a deformation
width of 14% (9.8 mm) in the x-direction, and the phantom was imaged again. An axial
and a coronal T1-weighted fast multi planar gradient echo (FMPGR) sequences were
performed in the uncompressed and compressed case.
10.1.3. Material properties of silicon phantom

The elastic properties of the phantom materials were evaluated on an Instron Model
1331 (Cambridge, MA) mechanical testing machine containing a semi-hydraulic com-
puter driven system for very accurate tensile strength measurements. We used a load cell
rated to 10 Newtons. This load cell is certified to an accuracy of 0.01 N. Flat cylindri-
cal samples of the silicone gel and the stiffer inclusion underwent uniaxial stress tests.
Static load-deformation (stress–strain) curves were obtained.
The silicon gel can be assumed to be an isotropic hyperelastic material, and be de-
scribed by a “strain energy potential” which defines the strain energy stored in the ma-
terial per unit of reference volume. The experimental data was fit to a law of rubber-
like material known as the Mooney–Rivlin strain energy function (G REEN and Z ERNA
[1968], S PENCER [1980]). The form of the Mooney–Rivlin strain energy potential is
1 el 2
U = C10 (I 1 − 3) + C01 (I 2 − 3) + J −1 , (10.1)
D1
where U is the strain energy per unit of reference volume; C10 , C01 and D1 are
temperature-dependent material parameters; I 1 and I 2 are the first and second devi-
2 2 2 (−2) (−2) (−2)
atoric strain invariants defined as I 1 = λ1 + λ2 + λ3 and I 2 = λ1 + λ2 + λ3 ,
where the deviatoric stretches λi = J −1/3 λi ; J is the total volume ratio; J el is the elastic
volume ratio, and λi are the principal stretches (defined as the ratios of current length
to length in the original configuration in the principal directions of a material fiber).
We assume that the silicon gel is incompressible and temperature independent (in the
conditions of the experiment). Therefore, the strain energy potential expression can be
simplified to
U = C10 (I 1 − 3) + C01 (I 2 − 3). (10.2)
C10 and C01 are the material parameters to be determined experimentally.
The stress–strain relationship is developed using derivatives of the strain energy func-
tion with respect to the strain invariants. We define this relation in terms of the nominal
stress TU (the force divided by the original, undeformed area) and the nominal, or en-
gineering strain εU (ratio of deformation length to length in the original configuration).
The deformation gradient, expressed in the principal directions of stretch is

λ1 0 0
F = 0 λ2 0 , (10.3)
0 0 λ3
where λ1 , λ2 and λ3 are the principal stretches. Because we assume incompressibility
and isothermal response, J = det(F) = 1 and, hence, λ1 λ2 λ3 = 1. The deviatoric strain
invariants in terms of the principal stretches are then I1 = λ21 + λ22 + λ23 and I2 = λ−2
1 +
λ−2
2 + λ−2
3 . The uniaxial deformation mode is characterized in terms of the principal
stretches, λi , as

λ1 = λU , λ2 = λ3 = 1/ λU , (10.4)
where λU is the stretch in the loading direction. The strain energy potential expression
can therefore be expressed solely in terms of λU :
−2
U = C10 λ2U + 2λ−1U − 3 + C01 λU + 2λU − 3 . (10.5)
To derive the uniaxial nominal stress TU , we invoke the principle of virtual work (δU =
TU δλU ) so that
∂U
TU = = 2 1 − λ−3U · λU C10 + C01 . (10.6)
∂λU
Now since the stretch λU is related to the nominal strain εU by (λU = εU + 1), the
nominal stress–strain relationship can finally be written as

TU = 2 1 − (1 + εU )−3 · (1 + εU )C10 + C01 . (10.7)
Eq. (10.7) was fit to the experimental stress–strain curves for the two types of silicon
gel, using the least sum of squares method. The results are shown in Fig. 10.2. The aver-
age parameter values calculated are C10 = 3740 ± 64 N/m2 , C01 = 1970 ± 34 N/m2 for
F IG . 10.2. Least-squares fit between the average experimental stress–strain curves (shown as dots), and the
Mooney–Rivlin model (shown as a continuous line), for the surrounding silicon gel material (top), and the
stiff inclusion gel material (bottom).
the surrounding silicone gel, and C10 = 16 300 ± 815 N/m2 , C01 = 10 490 ± 524 N/m2
for the silicone gel inclusion.
10.1.4. Phantom models

The first model was built using the BreastView software (see Fig. 10.1, bottom) (A ZAR ,
M ETAXAS , M ILLER and S CHNALL [2000]). The other model of the phantom was built
directly from its physical dimensions using a pre-processor program MSC/PATRAN

(MSC, CA), which automatically meshed the model. Both FEMs consist of 21 slices
(each 4 mm thick), stacked along the z-axis. The number of nodes in the x- and
y-directions is 18 and 22, respectively, in order to have square shaped volume elements.
The finite element models are made of 7497 elements each. The element material prop-
erties were given a homogeneous, isotropic, Mooney–Rivlin hyperelastic model, with
the C10 and C01 constants as measured above. The boundary conditions were applied
appropriately, and the 9.8 mm displacement of the pressure plate was modeled in the
initial conditions as a 9.8 mm displacement constraint on every node, which belongs to
the displaced surface of the phantom.
The finite element modeling simulations were done using:
(1) A robust finite element code ABAQUS/S TANDARD V.5.8 [1998], commercially
available. Each element was modeled as a hybrid incompressible solid 8-node
brick, which allows a fully incompressible constraint at each material calculation
point (ABAQUS/S TANDARD V.5.8 [1998]).
(2) The BreastView software in which 12 displacement iterations were used to com-
press the model. In each displacement iteration, a compression of 0.8167 mm
was applied to the relevant boundary nodes in the model. This represents an
average nominal strain of 1.16%, which can be considered small strain. The
maximum allowed error of integration eA in the adaptive algorithm, was cho-
sen to be (1/1000)th of the smallest dimension in the model dmin , which is
the smallest distance between two consecutive nodes: dmin = 3.90 mm, and
eA = 3.90 × 10−3 mm. This ensured full convergence of the solution after each
displacement iteration.
10.2. Results and discussion
The axial slice going through the center of the inclusion is shown in Fig. 10.3 (top) in
the uncompressed and in the compressed mode. As expected the edges of the phantom
have changed shape as well as the edges of the tumor. Because silicon is incompressible,
the side deformations of the phantom are quite large.
Because it is important in the real case to track the displacement of a cancer tumor
in the breast, we tracked the displacement of the inclusion in the phantom. By using
image analysis software, we measured the displacement vectors of the center of the
inclusion, as well as its eight corners. We used the axial slices to measure the x and y
displacements, and the coronal slices to measure the z displacements (the positions of
the various points of interest were calculated from the MR images with respect to the
position of the non-moving bottom edge of the silicon phantom).
10.2.1. ABAQUS simulation results

Fig. 10.3 (bottom) shows the displacement vectors of the inclusion corners and center
from MR data (bottom left), and from the ABAQUS model (bottom right). The aver-
age errors in displacements were 0.34, 0.66 and 0.40 mm in the x, y and z directions,
respectively, and are within the maximum imaging error.
Methods for modeling and predicting mechanical deformations
F IG . 10.3. Uncompressed and compressed axial MR slice of phantom (top left), 3D view of model including axial slice through center of inclusion and axial view
of inclusion center, before and after compression (top right), displacement vectors of inclusion corners and center from MR data (bottom left), and from the ABAQUS
model (bottom right).
613
10.2.2. BreastView simulation results

Fig. 10.4 (top) shows a comparison between the BreastView and ABAQUS simulated
virtual compressions. The displacement errors between the BreastView, ABAQUS sim-
ulations and the MR results are shown in Fig. 10.4 (bottom). The results show that the
displacement errors between the two simulations and the MR data are very close and
within the imaging error magnitude. Furthermore, all of the average displacement errors
per node between the two simulations are under 1 mm.
This silicon phantom study shows that:
• the compressed model allows us to track the position and motion of the stiff inclu-
sion in the real compressed deformable environment,
• after comparing the compressed BreastView and ABAQUS model results, using a
small strain approximation instead of the Lagrangian finite strain expression, in
our finite element algorithm does not introduce a significant error in simulations of
large deformation.
Although in this phantom study, the compressed model allows us to track the position
and motion of the stiff inclusion in the real compressed deformable environment, this
result does not necessarily extend flawlessly to an actual tissue model: patient breast
compression experiments are still needed for confirmation of our mathematical model
applied to human tissue deformation.
F IG . 10.4. Axial slice going through the inclusion after compression, in the ABAQUS simulation (left
model), and in the BreastView simulation (right model); displacement errors of inclusion corners and cen-
ter: ABAQUS–MR (left diagram), BreastView–MR (right diagram).
11. A patient study
A healthy patient’s left breast was used to track the displacement of specific landmarks
under plate compression. It is usually extremely difficult to accurately and locally track
the displacement of tissue in the breast before and after deformation. One may only
reliably quantify the displacement of clearly identifiable structures inside the breast, es-
pecially if these structures are small and/or rigid enough. We found a small “point-like”
cyst inside the breast, which was used as the inner landmark for tracking displacements
inside the breast after compression. Additionally, two vitamin E pills, used as land-
marks, were taped on the superior and inferior parts of the breast, in order to help track
the movement of the breast (vitamin E pills appear as bright spots on the MR images).
11.1.1. Patient breast MR data acquisition

The breast was imaged using a 1.5 T machine Signa Horizon Echospeed (GEMS, Mil-
waukee). 3D image sets were obtained under plate pressure conditions. The entire breast
was imaged medially to laterally, from the visible rib cage to the nipple. The MR ac-
quisition sequence was a 3D fast SPGR (T1 weighted), with a TR of 11.3 ms, a TE of
4.2 ms and a 25 degrees flip angle. We used a phased array multi coil. 124 slices each
with a 0.9 mm thickness were acquired sagittally (left to right). Each slice had a field of
view of 230 × 230 mm (256 × 256 pixels). This amounts to having cubic voxels with
0.9 mm sides.
11.1.2. Displacements due to plate compression

Usually, two compression plates are used to compress the breast (Table 11.1 shows the
plate to plate distances).
We make the following assumptions, which correspond to the actual breast compres-
sion setup (see Fig. 10.1, top, and Fig. 11.1):
– the plates are parallel to the y-axis;
– the plates move in the direction of the x-axis, towards the breast, in a way that
displacements can be considered static;
– the total compression distance Dtot for a plate is from its first point of contact with
the breast to its final resting position.
Once the physical parameters are known, such as the plate length Lp , the distance
from its tip to the patient’s rib cage Dt , and the total compression distance Dtot , we
TABLE 11.1
Plate to plate distances before and after compression
Plate to plate distance Compression amount

(mm) (mm)
Uncompressed 64.8 ± 0.9 0.0

Compressed 47.7 ± 0.9 17.1 ± 1.8
F IG . 11.1. Axial slice of model showing initial displacement parameters.
translate those parameters into prescribed displacements to all individual model surface
nodes which come in contact with the plate.
The total compression distance is translated into mtot number of displacement itera-
tions. A simple collision detection algorithm determines which node i comes in contact
with the plate at each displacement iteration. Each of these nodes is assigned a pre-
scribed displacement increment dni repeated in every iteration for which the plate is
in contact with the node. Therefore, the number of iterations for every surface node i
varies and is denoted mni .
11.1.3. Boundary conditions

Let Xi = (xi , yi , zi ) be the position of node i in the model (Xi is also a function of
time t), qi = (qi,x , qi,y , qi,z ) represent the displacement degrees of freedom for every
node i in the model. The boundary conditions are applied to the displacement degrees
of freedom as follows:
• Base of the breast. All nodes in the first two planes at the base of the breast model
are fixed, and represent the patient rib cage area. This provides the support structure
and fixes the breast model in space. Mathematically this is represented by

qi,y = 0,
∀ nodes i ∈ {base of the breast},
qi,z = 0.
• Interface between breast and rib cage. The nodes which are part of the breast tissue
in contact with the rib cage, are allowed to slide against the nodes which represent
the rib cage. Mathematically this is represented by
∀ nodes i ∈ {breast tissue at the base of the breast},

yi yRib cage,i (t = 0).
• Contact with the lateral plate. Boundary conditions between the virtual plate and
the breast model could cause increased displacement errors for the nodes close to
the edges of the virtual plate due to possible large rotational effects at the edges.
However, since we are concerned with lesion displacement, which is most gener-
ally located far enough from the plate edges, we can neglect errors due to rotations
in the vicinity of the plate’s edges.
Additional experiments were done in order to determine whether the breast skin
slides against the compression plates, or whether it sticks to them. Three different
patient breasts were compressed and imaged using the same protocol as in inter-
ventional procedures. A vitamin E pill was attached to the skin of the patient breast
directly in contact with the compression plate. The total average displacement for
the vitamin E pills on all breasts (compressed–uncompressed) was 8.8 ± 1.3 mm
(x = 7.8 ± 0.9 mm; y = 4.2 ± 0.9 mm). The results show that there is a sliding
effect between the skin and the compression plate. We approximate the sliding
by allowing the boundary nodes in the model directly in contact with the virtual
compression plate, to slide against the plate. Mathematically this is represented by
∀ nodes i ∈ {lateral side of breast},
if (node i in contact with lateral plate) then { qi,x = dni ,
where dni is the prescribed displacement at every contact iteration for node i.
• Contact with the medial plate. In principle, the patient breast is supposed to be
pushed against the medial compression plate as much as possible, in order to in-
crease contact with the plate, and decrease motion artifacts when using the lateral
plate. In practice, the initial prone lying position of the patient determines how
close the breast is to the medial plate; it is quite common that the breast is not com-
pletely pushed against the medial plate, and accounts for uncertainties in terms of
the contact and boundary conditions between the breast and the medial plate.
In order to minimize the uncertainty and possible sources of error that may result,
we model a real one-plate compression, with a virtual two-plate compression. We
suppose that the left plate does all the compression and the right plate remains im-
mobile. Then instead of applying the full compression amount using only the left
virtual plate as would be expected, we divide that compression amount evenly, and
both left and right virtual plates apply each half of the compression amount (since
the breast is compliant, we assume that pressure is equally distributed on each side
of the breast). However, after every compression step, the entire breast model is
translated back a distance equal to half of the total compression amount, parallel
and in the direction of the immobile right plate. In this way, it appears that the
model is compressed using only one moving virtual plate. We will compare that
method to the one-plate virtual compression, and choose the method, which yields
the best results. Mathematically this is represented by
∀ nodes i ∈ {medial side of breast},
if (node i in contact with medial plate) then { qi,x = dni ,
where dni is the prescribed displacement at every contact iteration for node i.
The breast was imaged first uncompressed, then imaged under plate compression. The
breast compression device, made by GEMS, stabilized the patient’s breast well enough
to minimize motion artifacts between image sets. The compression plates were able to
slide perfectly in the sagittal direction and were locked in position when the desired
compression was achieved. The displacements can therefore be considered static.
In order to help track the movement of the breast, we taped to the surface of the
breast two vitamin E pills (one towards the superior part, the other one towards the
inferior part) which appear as bright spots on the MR images: these serve as landmarks.
The pills were taped to the breast in such a way that they do not come into contact
with the compression plates, and did not influence the boundary conditions between
the compression plates and the breast skin. Gadolinium pills were embedded inside the
compression plates. These also appear as bright spots on the MR images, and allowed
us to confirm the compression distance between the two plates. The right plate (on the
breast’s medial side) was kept immobile, while the left plate (on the lateral side) was
moved to compress the breast.
11.1.4. Patient breast deformable model

A model of the patient breast was constructed from the given MR data, the boundary
conditions, and the applied displacements, using BreastView. The different breast tissues
inside the breast were modeled. The breast data was segmented semi-automatically,
starting at the intersection of the outer side of one compression plate and the breast, all
the way to the other plate.
The deformable model of the breast was built using every other slice out of the ex-
perimental data set, and discarding the outer slices, which did not contain useful in-
formation. A total of 58 slices were used. The breast model consists therefore of 58
slices (each slice in the uv-plane) stacked up along the z-axis. There are 8 nodes in
the x-direction and 8 nodes in the y-direction in every slice. The full model contains
3712 nodes, 2793 3D hexahedral elements, 2394 2D triangular elements, and is shown
in Fig. 11.2.
F IG . 11.2. Full breast model.

11.1.5. Mechanical measurements of human breast tissue and skin

Fig. 11.3 (right) shows the major structures of a typical mature pre-menopausal breast.
The dimensions and weight of the breast can greatly vary per individual. The breast is
an inhomogeneous structure containing many layers of many different kinds of tissue.
However, the two predominant types of tissue within the breast are fat and normal glan-
F IG . 11.3. Example of fatty tissue compartmentalization (left), and structure and location of Cooper’s liga-
ments (right).
dular tissue, which supports lactation. The mammary gland forms a cone with its base
at the chest wall and its apex at the nipple. Normal skin thickness lies between 0.5 and
1 mm. The superficial layer (fascia) is separated from the skin by 0.5 to 2.5 cm of sub-
cutaneous fat. Tentacle-like prolongations of fibrous tissue extend in all directions from
this fascia to the skin; these are called Cooper’s ligaments. In the adult mammary gland,
there are 15 to 20 irregular lobes, converging to the nipple through ducts 2 to 4.5 mm in
diameter (E GAN [1988a]). These ducts are immediately surrounded by dense connec-
tive tissue, which acts as a supporting framework. The glandular tissue is supported by
estrogen; when a woman reaches menopause the estrogen levels decrease and the glan-
dular tissue atrophies and eventually disappears, leaving only fat and skin (H ARRIS ,
L IPPMAN , M ORROW and H ELLMAN [1996]).
Carcinomas in affected breasts are usually accompanied by local changes in mate-
rial properties, due to increased stiffness in the lesion, as well as its shape and size
(E GAN [1988b]). There are several types of breast cancer, however the most com-
mon are ductal carcinoma (which begins in the lining of the milk ducts of the breast),
and lobular carcinoma (which begins in the lobules where breast milk is produced)
(NCI [1998]).
It has been observed that mechanical properties of soft tissues are due to their struc-
ture rather than to the relative amount of their constituents (F UNG [1987]). Whatever the
technique adopted, in most cases, the mechanical properties of soft tissues in vivo can
hardly be completely assessed. The constants found in one test only apply to the state
of deformation (strain rate, strain range involved (K ROUSKOP, W HEELER , K ALLEL ,
G ARRA and H ALL [1998])), which provides these constants, and no normal response
can be uniquely defined for the material. When simple relationships can be written, it
is generally only for limited ranges of stresses and strains. Furthermore, the mechanical
behavior derived from experiments cannot be readily correlated with in vivo conditions
of the tissue where generally the reference state is not completely known (L EE and
T SENG [1982]). The assumptions leading to the results of the experiments may there-
fore no longer be valid in the real conditions of functioning (C RISP [1972]). Experi-
ments are usually limited to a one-dimensional stress field (YAMADA [1970] presents
a relatively broad range of data; much of the data are derived from animal experiments
and all the information relates to uniaxial tensile properties of soft tissue). It is clear
that all stresses and strains in the three dimensions are involved, and should be consid-
ered for an accurate characterization of breast tissue. But this would require extensive
two- or three-dimensional testing programs, which are difficult to carry out on small
specimens that must be maintained in living conditions. It is thus difficult to assert the
absolute significance of the material models developed (F UNG [1972]). This is why only
average values of experimentally derived material models are used in our finite element
analysis.
Few studies have been made on determining the mechanical properties of tissue in the
breast, however average values of Young’s modulus have been calculated for fat, glandu-
lar tissue, and cancer tissue (S KOVORODA , K LISHKO , G USAKYAN , M AYEVSKII , Y ER -
MILOVA , O RANSKAYA and S ARVAZYAN [1995], K ROUSKOP, W HEELER , K ALLEL ,
G ARRA and H ALL [1998], L AWRENCE , ROSSMAN , M AHOWALD , M ANDUCA , H ART-
MANN and E HMAN [1999]). Since we need non-linear stress–strain curves describing
the mechanical behavior of breast tissue, we will use the mechanical properties of breast
tissue determined by W ELLMAN and H OWE [1998]. Exponential curves (which have
been used in the past for several different tissue types (F UNG [1993])) are used to de-
scribe the stress–strain properties of breast tissue, following experimental stress–strain
curves obtained from uniaxial loading of breast tissue. The fresh tissue samples were
tested in the operating room within 10 minutes of excision, were kept hydrated by pe-
riodic application of saline solution, and were tested at room temperature (21 ± 2.5 ◦ C)
(W ELLMAN and H OWE [1998], W ELLMAN [1999]). The exponential curves describing
the elastic modulus En for tissue type n, are given by
∂σn
En = = b · emεn , (11.1)
∂εn
where σn and εn are the nominal stress and strain, respectively, for tissue type n. b and
m are the fit parameters determined experimentally for every tissue type:
• bglandular = 15 100 N/m2 ;
• mglandular = 10.0 (within 1 standard deviation);
• bfat = 4460 N/m2 ;
• mfat = 7.4 (within 1 standard deviation).
We used a value of 0.49999 for Poisson’s ratio.
The mechanical properties of skin have been studied more thoroughly than those
of breast tissue, and several papers have been written on the subject (V ERONDA and
W ESTMANN [1970], AGACHE , M ONNEUR , L EVEQUE and D E R IGAL [1980], F UNG
[1981], S CHNEIDER , DAVIDSON and NAHUM [1984]). Though it is not a homogeneous
material, in many cases skin can be simplified to be statistically homogeneous (at higher
strains it was found that the slope in the linear region of the stress–strain curve is similar
in all directions) (M AUREL , W U , M AGNENAT T HALMANN and T HALMANN [1998]).
The typical experimental stress–strain curve for skin (E LDEN [1977]) is transformed
into a piecewise linear stress–strain curve which we will use to describe the mechanical
properties of skin in the breast model (as shown in Fig. 11.4).
F IG . 11.4. Piecewise linear stress–strain curve for skin (E LDEN [1977]).

TABLE 11.2
Parameters for skin material model (in N/m2 )
Parameter name Line segment 1 Line segment 2 Line segment 3
ai (N/m2 ) 3.43 × 106 2.89 × 107 1.57 × 108

bi (N/m2 ) 0 −1.36 × 107 −1.01 × 108
Valid strain range 0 ε 0.54 0.54 < ε 0.68 0.68 < ε 1
The stress–strain relationship for skin is of the form
σskin = ai · εskin + bi , (11.2)

where (ai , bi ) are the parameters for every linear segment of the curve (i = 1 to 3). Then
the elasticity modulus is given by
Eskin = ai (i = 1 to 3). (11.3)

The parameter values are shown in Table 11.2.
11.1.6. Modifications of fat tissue model

The rationale underlying the modification of the fatty tissue model is based on the fact
that ex-vivo measurements of fatty tissue properties, although accurate, always ignore
the supporting structure of fibers in the breast. Also those measurements are almost
always made at room temperature, which significantly affects the mechanical proper-
ties of fatty tissue, since fatty tissue is almost liquid at body temperature (K ROUSKOP,
W HEELER , K ALLEL , G ARRA and H ALL [1998]) (on average 10–15 ◦ C higher than
room temperature).
None of the experimental measurements of breast tissue encountered in the literature
take into account the supporting structure of fibers including Cooper’s ligaments to de-
scribe the mechanical behavior of fatty tissue in vivo. We also know from the literature,
and from testing with other tissue types, how the absolute value of the tissue modulus
may be affected by the boundary conditions (confinement) acting on a sample. Confine-
ment can be significantly different from organ to organ depending on its surrounding
environment (K ROUSKOP, W HEELER , K ALLEL , G ARRA and H ALL [1998]).
We hypothesize that:
• The supporting structure of fibers, including Cooper’s ligaments, compartmental-
izes fatty tissue, and prevents it from being squeezed out of its location.
• As fatty tissue is being compressed, the local pressure increases and leads to an
increase in the apparent stiffness value of fat.
These hypotheses are supported by the experimental evidence of the numerous com-
pressed patient breast images scanned to date using MRI. This data shows fatty tissue
clearly not squeezed out but actually remaining in its location as the breast is being
compressed, as shown in Fig. 11.3.
We test the hypotheses by updating the material model, such that the stiffness of
fatty tissue is made to increase up to an average stiffness value of glandular tissue:
as fatty tissue is being compressed and compartmentalized, the local pressure in the
compartment increases and leads to an increase in the apparent stiffness value of fat,
until the stiffness reaches that of the glandular tissue. Given this hypothesis, we can set
the following necessary boundary conditions for Efat :
• If εfat = 0: Efat (0) = bfat ,
∂Efat
• If εfat εlimit: bfat Efat Egland, 0,
∂εfat
• If εfat = εlimit: Efat = Egland = bglandemgland ·εlimit ≡ α, (11.4)
∂Efat ∂Egland
= = bglandmglandemgland ·εlimit ≡ β,
∂εfat ∂εgland
• If εfat εlimit: Efat = Egland.
These boundary conditions merely state the fact that Efat is a continuous smooth
non-decreasing function.
The simplest equation, which satisfies the above conditions, is a quadratic equation
of the form
Efat (εfat ) = A · εfat

2
+ B · εfat + C. (11.5)
Solving for A, B and C given the boundary conditions, yields the following:
β · εlimit + bfat − α 2α − 2bfat − β · εlimit
A= 2
, B= , C = bfat . (11.6)
εlimit εlimit
There is a condition on εlimit which comes from the necessity that (Efat ) 0; this is
equivalent to writing (Efat ) (0) 0 as long as A > 0. This leads to the condition that
εlimit (2 · (α − bfat ))/β. The curve describing Efat (εfat ) is shown in Fig. 11.5 using the
experimentally derived values for bgland, mgland and bfat . We solved for the maximum
allowed strain limit, and used it to model fatty tissue. We find εlimit = 15.5%.
We test the updated fatty tissue model hypothesis, and the value of εlimit , in the sen-
sitivity analysis, where the effect of varying the material parameters on the model per-
F IG . 11.5. Fat material properties curve.

formance is investigated. Rather than modeling the structure and geometry of Cooper’s
ligaments, we model their functionality and overall effect.
11.2.1. MR results
Table 11.3 shows the displacement vectors of the cyst inside the breast, and the two
vitamin E pills after compression. The displacement measurements (in mm) were made
using Scion Image software. In order to compute the cyst and vitamin E pills dis-
placements, we calculated the positions of the breast cyst and vitamin E pills with
respect to the position of the gadolinium pills embedded in the non-moving compres-
sion plate. The edge points of the vitamin E pills closest to the skin were used in the
calculations. The images in Fig. 11.6 show the axial cross-sections of the patient’s left
breast.
11.2.2. Variational study

The effect of varying different parameters is investigated. Several parameters describ-
ing the model are varied over a physiologically relevant range, and every time the model
simulation is done, the displacement of the landmarks, as well as other relevant perfor-
mance assessment parameters are recorded.
TABLE 11.3
Displacement vectors (mm) of the landmarks obtained experimentally (compressed–uncompressed)
Vitamin E pill Vitamin E pill Small cyst

(superior) (inferior)
X Y Z X Y Z X Y Z
Compression −9.0 3.6 2.7 −7.2 0.9 −3.6 −6.3 1.8 1.8
F IG . 11.6. MR slices containing the landmarks, for the uncompressed and compressed cases.
First, we decide which of the one-plate or two-plate virtual compression yields the
best results. Then after a convergence analysis we decide whether the slab model (vari-
able z-mesh density) can replace the full model in the virtual compression; this would
dramatically decrease the simulation time. Finally, we do a material properties sensi-
tivity analysis to show how sensitive is the model performance to variations in material
parameters.
11.2.3. Performance assessment parameters

Two performance assessment parameters are used in the variational study:
• Displacement difference (DISPdiff ): difference (model–real) of displacement, of
inclusion center of gravity, where inclusion can be lesion, cyst, or vitamin E pill,
2
DISPdiff = Xdiff + Ydiff
2 + Z2
diff . (11.7)
• %Misclassification (%MIS): compares the number of misclassified pixels in the
model with the compressed MRI,
%MIS = NMIS/TNUM, (11.8)
where NMIS is the number of misclassified pixels in model with respect to com-
pressed MR dataset ((nonfat pixels in fat elements) + (nonglandular pixels in glan-
dular elements) + (nontumor pixels in tumor elements)), and TNUM is the total
number of breast pixels in the MR dataset.
The displacement differences are used to assess the performance of the most impor-
tant result in the model, which is how well the displacement of an inclusion in the real
breast, can be predicted using the model. The misclassification percentage is a value,
which gives an idea of how well the model globally predicts the displacement of the
major structures in the breast.
Therefore, the displacement difference offers a local measure of performance,
whereas the misclassification percentage offers a global “macroscopic” measure of the
performance of the breast model.
11.2.4. Initial uncompressed model and uncompressed MR images

The uncompressed MR images containing the landmarks and the corresponding un-
compressed model slices are shown in Fig. 11.7. The arrows in the figure indicate the
location of the landmarks.
All of the simulations in the variational studies were done on a SGI Octane Worksta-
tion with 2 195 MHz IP30 processors (MIPS R10000 processors), and 256 Megabytes
of memory (RAM). The first step in the variational analysis is to decide what type of
model to use, and what type of virtual compression to apply. Given the different pa-
rameters outlined above, the virtual compression was done using one then two virtual
compression plates, with the full model and the slab model. The displacement differ-
ences and %misclassifications are shown in Fig. 11.8.
11.2.5. One-plate vs. two-plate virtual compressions for the full model
The full model consists of 58 slices stacked up along the axial axis. Each slice contains
8 × 8 nodes, for a total of 3712 nodes, 2793 3D hexahedral elements, and 2394 2D tri-
F IG . 11.7. Uncompressed MR images (top), and uncompressed model slices (bottom).
F IG . 11.8. Displacement differences (mm) and %misclassifications in the one-plate virtual compression
(left), and in the two-plate virtual compression (right), with the full model.
angular elements. This experiment is aimed at comparing the performance of the virtual
one-plate compression simulation to that of the virtual two-plate compression simu-
lation, where the total compression amount is evenly divided between the two virtual
compression plates.
Clearly, from the results shown above, the two-plate virtual compression yielded bet-
ter results than the one-plate virtual compression, using the full model.
The %misclassifications for the slices containing the vitamin E pills are high com-
pared to that of the slice containing the cyst. This higher level of inaccuracy is due to the
fact that the former slices belong to the outer edges of the patient’s breast; the contact
and boundary conditions between these slices and the compression plates is not always
clear, and it is very difficult to accurately predict their deformation for that reason.
11.2.6. Convergence analysis

We did a convergence analysis in the (x, y) direction: the (x, y) mesh density was in-
creased, and the model compression was done. We recorded the total displacement of
F IG . 11.9. Convergence analysis in the (x, y) direction (top), convergence analysis in the z-direction (mid-
dle), and convergence analysis with the number of displacement iterations per simulation step (bottom). The
density factor on the top two graphs represents the total number of nodes in the model, normalized by the
number of nodes in the model with 8 × 8 nodes in the (x, y) direction.
the cyst. Another convergence analysis was done in the z direction, using the slab model:
the z-direction mesh density was varied using the slab-model algorithm, increasing the
number of model slices around the point of interest. We recorded the total displacement
of the cyst. The number of displacement iterations was increased (as the displacement
per iteration was decreased in order to keep the same total virtual plate compression of
17 mm). The analysis was done on the displacement of the cyst in the model. All the
results are shown in Fig. 11.9.
The analysis shows that the simulation indeed converges to the solution, as the model
mesh is refined. Also using about 10 displacement iterations per simulation step proves
enough for convergence of the solution. Furthermore, using the slab model with only
8 slices instead of 56 does not decrease the performance of the simulation. In order to
show the performance of the slab model, we ran the entire simulation using that model.
The slab model contains 512 nodes, 343 3D hexahedral elements, and 294 2D triangular
elements. As shown in Fig. 11.10, the mesh density in the slab model is the highest
around the points of interest (cyst, Fig. 11.10, center or vitamin E pill, Fig. 11.10, right).
The results from the simulation are shown below.
F IG . 11.10. Full breast model mesh (top), and variable meshes around different points of interest (cyst,
center and vitamin E pill, bottom).
(a)
(b)
(c)
(d)
F IG . 11.11. Displacement differences (a), and %misclassification (c) in the two-plate virtual compression
with the slab model; compressed MR slices (b), and corresponding virtually compressed model slices (d)
after a two-plate virtual compression using the slab model.
Cyst Vit. E pill (INF) Vit. E pill (SUP)

Disp. difference 1.3 (0.1) 3.5 (0.1) 4.6 (0.1)
%Misclassification 11.8 (1.0) 25.1 (1.2) 31.4 (0.5)
F IG . 11.12. Graph showing sensitivity analysis results (top), and table showing averages and standard de-
viations (in parentheses) of sensitivity analysis (bottom). Each point in the sensitivity analysis represents a
modeling experiment with one of the material parameters varied. We can clearly distinguish three different
clusters. The lower left cluster represents results from the cyst displacement, the middle cluster represents
results from the vitamin E pill (inferior) displacements, and the upper right cluster represents results from the
vitamin E pill (superior) displacements.
When comparing the results in Fig. 11.11 (top) to Fig. 11.8, we see that the perfor-
mance of the slab model was as good as that of the full model. This enables us to run
the simulation dramatically faster since the slab model contains many less elements than
the full model.
11.2.7. Material properties sensitivity analysis

The effects of varying the different material properties parameters are investigated. The
parameters are varied over a physiologically relevant range, within two standard devia-
tions of their average experimentally derived value. The numerical values of the relevant
material properties used in our chosen model were varied within ±1.2 standard devia-
tions, one at a time. The parameter for the new updated fat model was varied by 30%,
then by 60%. The material parameters of interest which were varied are mgland , bgland for
the glandular material model, and elimit for the new updated fat model. All of the results
are shown in Fig. 11.12, where the displacement differences vs. %misclassifications are
graphed. Each point represents a modeling experiment with one of the material para-
meters varied. We can clearly distinguish three different clusters. The lower left cluster
represents results from the cyst displacement, the middle cluster represents results from
the vitamin E pill (inferior) displacements, and the upper right cluster represents results
from the vitamin E pill (superior) displacements.
12. A clinical breast compression study
Three patient breasts undergoing breast MRI were imaged uncompressed and then com-
pressed. Patients were chosen to have a variety of breast cancers of different shape, size
and location in the breast. Patient ages vary between 49 and 71 years of age. In all cases
the breast affected is the left breast. Subsequent biopsy revealed cancer in all cases.
12.1.1. Initial and boundary conditions

The breasts were imaged first uncompressed, then imaged under plate compression, as
the patients were lying prone. The right compression plate (on the breast’s medial side)
was kept immobile, while the left plate (on the lateral side) was moved to compress the
breast. Table 12.1 shows the compression amounts.
12.1.2. Patient breast deformable models

A model of each patient’s breast was constructed from the given MR data, the boundary
conditions, and the applied displacements, using BreastView. The different breast tissues
inside the breast were modeled. The breast data was segmented semi-automatically,
starting at the intersection of the outer side of one compression plate and the breast, all
the way to the other plate. The breast models consist of 8 parallel slices stacked along
the z-direction, distributed in such a way as to maximize the number of model slices
around the tumor (using the variable meshing algorithm). The number of nodes in the
plane of every slice (x, y plane) was chosen as to allow the breast tumor to be fully
included inside one 3D element, but at the same time allowing the simulation to run in
less than 15 minutes, as shown in Table 12.2.
The displacements of the tumors’ center of gravity were recorded both in the real
breasts and in the deformable models. All of the simulations were done on a SGI Oc-
tane Workstation with 2 195 MHz IP30 processors (MIPS R10000 processors), and 256
Megabytes of memory (RAM).
TABLE 12.1
Breast compression amounts
Patient 1 Patient 2 Patient 3
Compression amount (mm) 15.3 ± 1.8 14.4 ± 1.8 34.2 ± 1.8

% Compression 24 22 22
Average size of lesion (mm) 7.2 5 13
TABLE 12.2
Breast deformable models
(x, y) number of nodes (10, 12) (10, 10) (8, 12)

Total number of nodes 960 800 768
Number of 3D elements 693 567 539
Fig. 12.1 shows the uncompressed MR slices for the three patients, containing the le-
sions indicated by an arrow (top), and uncompressed corresponding model slices con-
taining the lesion elements (bottom). Fig. 12.2 shows the compressed MR slices for
the three patients, containing the lesions indicated by an arrow (top), and compressed
corresponding model slices containing the lesion elements (bottom).
F IG . 12.1. Uncompressed MR slices for the three patients, containing the lesion indicated by an arrow (top),
and uncompressed corresponding model slices containing the lesion element (bottom).
F IG . 12.2. Compressed MR slices for the three patients, containing the lesion indicated by an arrow (top),
and compressed corresponding model slices containing the lesion element (bottom).
Table 12.3 shows the experimental lesion displacements, the modeled lesion dis-
placements, and the displacement
differences (model–real) DISPdiff of inclusions cen-
ter of gravity, DISPdiff = ( xdiff2 + y 2 + z2 ), where x
diff diff diff = xmodel − xreal , ydiff =
ymodel − yreal and zdiff = zmodel − zreal . (xmodel, ymodel , zmodel ) and (xreal , yreal , zreal ) are
the displacement vectors (compressed–uncompressed) of the inclusions center of grav-
ity obtained respectively from the model, and experimentally.
TABLE 12.3
Displacement vectors (mm) of the lesion obtained experimentally, and from the model (compressed–
uncompressed), displacement differences DISPdiff , both in mm and as a percentage of lesion size
Lesion displacement (mm) x y z x y z x y z

– experimental −5.4 1.8 −2.7 −10.8 0.9 0.0 −18.9 5.0 3.6
– model −4.6 0.3 −0.6 −10.1 0.3 0.5 −14.0 0.1 1.3
DISPdiff (mm) 2.7 1.0 7.3
DISPdiff as a % of lesion size 37.5 20 56
13. Registration of breast MR images of the same patient under different

compressions
A healthy patient’s breast was imaged under two different plate compression amounts.
A deformable model of that breast was constructed and compressed using virtual com-
pression plates. The displacement of a small cyst was recorded in the real breast and
compared to the displacement of the ‘virtual cyst’ in the deformable model from one
compression state to the other.
13.1.1. Initial and boundary conditions

The breast was imaged under two different plate compression amounts: 12 and 26%,
resulting in a compression of 9.0 ± 1.8 mm between the two states. The right compres-
sion plate (on the breast’s medial side) was kept immobile, while the left plate (on the
lateral side) was moved to compress the breast.
13.1.2. Patient breast deformable model

A deformable model of the breast was constructed using the least compressed breast
MR data, and deformed using virtual compression plates. The breast model consists of
58 slices with 8 nodes in the x-direction, and 8 nodes in the y-direction in every slice.
The full model contains 3712 nodes, and 2793 3D elements.
The displacement of a small cyst was recorded both in the real breast and in the
deformable model from one compression state (12%) to the other (26%). The simulation
was done using the custom-written BreastView software. The simulation was done on a
SGI Octane Workstation with 2 195 MHz IP30 processors (MIPS R10000 processors),
and 256 Megabytes of memory (RAM).
13.2.1. Rotational differences in breast positioning

One of the major concerns to this experiment lies in the initial and boundary conditions.
In our method, registering two breast image sets of a same patient taken under different
pressure plate conditions and at different times, does not account for the possibility
of rotational differences in breast positioning. Translational differences are trivial to

account for, however, rotational differences may cause significant errors in the modeling
results if not taken into account.
In order to assess the amount of rotational differences in breast positioning from one
instance to another, four different patients from a previous study were used. The patient
breasts were scanned at two different times, weeks apart, and sagittal MR slices from
corresponding image sets, and showing the same location in the breast, were compared
to each other.
Results are shown below in Fig. 13.1. A qualitative look at the superimposed breast
images in Fig. 13.1(c) leads to the surprising conclusion that rotational differences due
F IG . 13.1. Sagittal cross-sections of four patient breasts taken at two different times ((a) time 1, (b) time 2)
weeks apart, and a transparent superposition of the two cross-sections, showing clearly the shift in boundaries
from one time to another (c).
to initial positioning of the breast in the compression device are virtually inexistent.
There are some translational differences, and differences in shape due to the difference
in breast compression amount between the two time points. However, the results clearly
show that we can neglect the potential sources of error in the simulation, originating in
rotation differences due to breast positioning from one time point to another.
13.2.2. MR and model results

The MR imaging results are shown in Fig. 13.2. Table 13.1 shows the experimental
vs. the modeled cyst displacement, and Fig. 13.3 contains the slab of the breast model
showing the slice which contains the cyst in the initial 12% compression state (left) and
in the final 26% compression state (right).
The three patient breast simulations above show that the displacement differences
from experiment to model are smaller than the actual size of the lesions concerned.
The simulations therefore show that the model can reasonably predict the displacement
of the lesions involved. Table 12.3 indicates also that the displacement differences as
a percentage of lesion size are of the order of 50%, which shows that the modeled
displacement of the lesion may allow half of the lesion size to overlap with the modeled
lesion size.
Finally, the model has been shown to predict the displacement of lesions in a patient
breast undergoing plate compression, for lesions of about 5 mm or more in size.
The model performance is as reliable as the parameter that is most sensitive to vari-
ations in the conditions of the experiment. In our model, it is shown, that the physical
representation of the breast is most sensitive in variations of the breast shape. It is also
F IG . 13.2. MR slice of the patient breast containing the cyst.
TABLE 13.1
Experimental vs. modeled cyst displacement
Patient breast
Cyst displacement (mm) x y z

– experimental −1.8 0.9 1.8
– model −2.5 1.7 1.4
Displacement difference (mm) 1.4
F IG . 13.3. Slab of breast model showing the slice which contains the cyst in the initial 12% compression
state (left) and in the final 26% compression state (right).
shown through a material properties sensitivity analysis, that the performance of the
model has modest dependence to variations in material properties within 1 to 2 stan-
dard deviations. Indeed, the shape and size of a patient breast influence the amount
of compression on the breast, and the boundary properties between the breast and the
compression plates.
Therefore, additional experiments would be needed in order to assess the reliability
of the model by choosing a number of patients with breasts of different sizes, leading
to different initial and boundary conditions for the model. We may want to initiate a
study involving 12 patients in three categories of breast size: small, medium and large,
with four patients in each category. In each category we may want to track lesion dis-
placements for lesions that are located either in the middle of the breast, or towards the
outer edges on an axial axis. The results of this study will enable us to determine the
relationship between the minimum size of a trackable lesion, the size of the breast, and
the location of the lesion in the breast.
It may also appear from the results shown in Table 13.1 that errors in the x, y and z
dimensions, respectively, are of the order of 20 to 50%, which seems very high. How-
ever, the correct way to look those results is to consider the displacement difference in
absolute terms. Indeed, since we are tracking the displacement of a point object, the
radius of error is given by the displacement difference, which is 1.4 mm. This error in-
dicates that if the cyst was the center of a lesion with a size of the order of 3 mm (twice
the radius of error), it would be possible to track that lesion from the model simulation.
14. Potential sources of error
Three types of discretization errors can occur: first, errors from the non-linear material
properties model, then errors from the time domain discretization (solving the dynamic
equations), and finally, errors from the finite element method.
14.1. Sources of error from material properties measurements
It is difficult to assert the absolute significance of the material models developed

(F UNG [1972]). Furthermore, the actual complexity of breast tissue (anisotropy, in-
homogeneity, the number and distribution of Cooper’s ligaments) prohibits us from
accurately calculating a 3D map of a patient’s breast tissue properties. This is why
only average values of experimentally derived material models are used in our finite
element analysis. The non-linear stress–strain curve describing the mechanical be-
havior of breast tissue, is discretized into a number of different segments, each cor-
responding to a displacement iteration. In order to test the effect of material prop-
erties model discretization, the number of displacement iterations was increased (as
the displacement per iteration was decreased in order to keep the same total virtual
plate compression). The analysis was done on the displacement of the cyst in the
model.
As shown in the convergence analysis with the number of displacement itera-
tions per simulation step (Fig. 11.9), increasing the number of displacement itera-
tions, which also means increasing the number of linear segments discretizing the
non-linear material stress–strain curve, leads to a convergence of the two important
result parameters in the model. However, the variation in the parameters do not ex-
ceed 15% for the %misclassification, and 14% in the displacement of the cyst. Once
again since only average values of experimentally derived material models are used
in the analysis, the accuracy involved in the stress–strain curve discretization is not
as important as the accuracy involved in approximating the shape of the breast in the
model.
The material properties sensitivity analysis (Fig. 11.12) shows that large variations in
material properties parameters (±1.2 standard deviations) do not significantly affect the
parameter results. This may be explained by the fact that the breast is under pressure:
since glandular and fatty tissue make up the majority of breast tissue, and since the fat
material model eventually becomes equivalent to the glandular tissue material model at
a certain level of strain, then material properties difference may become of secondary
importance in the simulations. If that is the case, then we may be able to simplify the
material properties models and thus increase the speed of the breast compression simu-
lation.
14.2. Sources of error from the time domain discretization
The time-discretized nodal equation is

qi,t +1 = qi,t + t · fi,internal . (14.1)
At every displacement iteration, we integrate Eq. (14.1) using the adaptive Euler tech-
nique (P RESS , T EUKOLSKY, V ETTERLING and F LANNERY [1992], D E V RIES [1994]),
where the time step t, varies according to the amount of integration error.
We can make the integration error sufficiently small to insure convergence of the
solution. The numerical integration ends when the difference between the sum of all
displacements from one time iteration to the next is less than a threshold (10−3% is
F IG . 14.1. Time convergence curve: % incremental displacement (log scale).
shown to be a good assumption):

n
n
difference = |qi,t +1 | − |qi,t | < threshold. (14.2)
i=1 i=1
Fig. 14.1 shows that a difference between two consecutive displacements, of less than
10−3 % is enough to claim convergence of the model.
14.2.1. Sources of error from the finite element method

Errors in the finite element method can be divided into two classes:
• Discretization errors resulting from geometric differences between the boundaries
of the model and its finite element approximation.
• Modeling errors, due to the difference between the true solution and its shape func-
tion representation.
Using smaller elements can reduce discretization errors – the errors tend to zero, as
the element size tends to zero. Shape function errors do not decrease as the element
size reduces and may thus prevent convergence to the exact solution or even cause
divergence. There are two main criteria required of the shape function to guarantee
convergence:
• Completeness. A complete polynomial of order at least p must be used for the
representation of the variable within an element, where p is the order of the highest
derivative of the variable appearing in the energy functional (in three dimensions a
complete polynomial of order p can be written as

l
f (x, y, z) = ar x i y j z k , i + j + k p,
r=1
where the number of terms in the polynomial is l = (p + 1)(p + 2)(p + 3)/6).

• Conformity. The elements must be conforming, that is, the representations of the
variable and its derivatives up to and including order p − 1 must be continuous
across interelement boundaries, where p is the order of the highest derivative ap-
pearing in the functional.
An important property of isoparametric elements is that they provide C 0 continu-
ity, and contain a complete linear polynomial in Cartesian coordinates (E RGATOUDIS ,
I RONS and Z IENKIEWICZ [1968]). Therefore, isoparametric elements satisfy the two
criteria required of the shape functions to guarantee convergence.
In summary, although the breast deformable model has been shown to reasonably
predict the displacement and deformation of cancer lesions under plate compression, its
performance could be improved by:
• Better characterizing local material properties: presently, only average values of
experimentally derived material models are used in our finite element analysis.
However, strictly speaking, every little piece of tissue inside a patient’s breast has
a unique set of material properties, depending on its structure and composition. The
only way to better characterize local material properties for every individual patient
is through an in vivo technique. This technique should provide live non-invasive
quantitative information on the non-linear deformation properties of every location
in a patient’s breast. Such a technique does not exist yet today, however research
towards this goal is underway, through tissue elastography techniques.
• Better defining the boundary and initial conditions: no matter how complex the
deformable model of the breast may be, it becomes relatively useless without an
accurate quantitative description of the breast’s physical interaction with its sur-
rounding environment, i.e., how is the patient positioned, how compressed is the
breast, what is the size of the compression plates, how much contact is there be-
tween the plates and the breast, what is the friction between the plates and the
breast skin, how are the plates moved, etc.
• Using a geometrically more accurate breast model: smaller model elements for ex-
ample, will decrease discretization errors by allowing a structurally more accurate
description of the different breast tissues.
15. Specific properties of reliability
The model performance is as reliable as the parameter that is most sensitive to varia-
tions in the conditions of the experiment. In our model, it is shown, that the physical
representation of the breast is most sensitive in variations of the breast shape. It is also
shown through the material properties sensitivity analysis, that the performance of the
model has modest dependence to variations in material properties within 1 to 2 standard
deviations.
Indeed, the shape and size of a patient breast influence the amount of compression on
the breast, and the boundary properties between the breast and the compression plates.
For example, when testing the performance of the two-plate virtual compression in the
slab model, the displacement differences in mm were 1.3, 3.6 and 4.5 for the cyst, in-
ferior vitamin E pill, and superior vitamin E pill, respectively. The %misclassifications
where 11, 27 and 32%. It is clear that the cyst’s displacement was predicted more accu-
rately than the displacement of the vitamin E pills. This difference in accuracy is due to
the fact that the vitamin E pills were placed on the outer edges of the patient’s breast,
where the contact and boundary conditions between the breast and the compression
plates are not always clearly defined, and it may become difficult to accurately predict
breast deformation at these locations.
Therefore, additional experiments would be needed in order to assess the reliability
of the model by choosing a number of patients with breasts of different sizes, leading to
different initial and boundary conditions for the model. One would have to measure the
boundary conditions through physical measurements of the breast placement configura-
tion.
16. Major novelties in the model
The major novelties in this model include:

• The updated fatty tissue material model, which takes into account the effect of fat
compartmentalization due to Cooper’s ligaments in the breast. We showed through
empirical evidence that fat compartmentalization occurs as the breast is being com-
pressed, and that the new updated fatty tissue material model takes that phenom-
enon into account, and performs better than the original fatty tissue model.
• The use of small displacement iterations while updating the tissues’ different mate-
rial properties, in order to model the non-linear behavior of tissue material models.
The silicon phantom study allowed us to compare predicted displacement of a sil-
icon inclusion from an ABAQUS (large commercially available FEM package)
simulation, and from a BreastView (our software package) simulation, and showed
that our algorithm does yield accurate results.
• The application of finite element modeling theory to model the deformation of a
human female breast in such a way that the entire process takes less than a half-
hour, which according to the clinicians consulted, is a reasonably short time dura-
tion.
17. Concluding remarks
Currently, High Field (1.5 T) Superconducting MR imaging does not allow live guid-
ance during needle breast procedures. The current procedure allows the physician only
to calculate approximately the location and extent of a cancerous tumor in the com-
pressed patient breast before inserting the needle. It can then become relatively uncer-
tain that the tissue specimen removed during the biopsy actually belongs to the lesion
of interest. A new method for guiding clinical breast biopsy was presented, based on a
deformable finite element model of the breast. The geometry of the model is constructed
from MR data, and its mechanical properties are modeled using a non-linear material
model. This method allows imaging the breast without or with mild compression before
the procedure, then compressing the breast and using the finite element model to predict
the tumor’s position during the procedure.
The final results show that it is possible to create a deformable model of the breast
based on the use of finite elements with non-linear material properties capable of mod-
eling and predicting the deformation of the breast. This study also shows that the full
procedure can be carried out in less than a half-hour: from start to end, the average times
to completion were 12 minutes for segmentation of MR data, 3 minutes for the model
mesh creation, and 14 minutes for the model simulation.
The results also suggest that it is possible to use the deformable model of the breast
in order to register lesion locations in image sets of the same patient breast taken at
different times, and under different pressure plate conditions. We showed qualitatively
that rotational differences in initial breast positioning are virtually inexistent, and can
therefore be neglected in the simulation.
This deformable model may be used as a new tool to the physician (A ZAR , M ETAXAS
and S CHNALL [2001]), who will:
image the breast under little or no compression (thus increasing the

1.
contrast and visibility of the tumor),
2. build the deformable model of the breast from the data,
compress the breast as much as the patient will allow (to minimize
3.
deformations caused by the insertion of the needle),
virtually compress the breast model in the virtual environment us-

4.
ing the same boundary conditions as in reality,
finally use the virtually compressed model to predict position of the

5.
real lesion within the real compressed breast during the procedure.
Future development in this project will be to develop a framework for a full-scale real-
time finite element simulation of breast tissue deformation in more complex systems
involving surgical instruments interacting with the model. The keys for such a devel-
opment will include scalable parallel solution algorithms (S ZEKELY, B RECHBUHLER ,
H UTTER , R HOMBERG and S CHMID [1998]), as well as dedicated parallel hardware.
The new system will allow real-time virtual surgical procedures of the breast, allowing
the physician or student to fully prepare or train for the real procedure. The real-time
capability will then allow the user to connect haptic devices such as a glove with pres-
sure feedback, or a simulated needle with force feedback, and will bring the user one
step closer to reality.
18. Appendix
18.1. Modeling 8-node hexahedral solid isoparametric elements
18.1.1. Finite element shape functions

The shape functions are used to interpolate a motion variable (displacement, position)
from the nodes to a point in the element. The shape functions for an eight-node brick
element (or linear solid element) are:
Ni = (1/8)(1 ± ξ )(1 ± η)(1 ± ζ ) (18.1)

in which i = 1, 2, . . . , 8.
18.1.2. Finite element stiffness matrix

The derivation of the 3D finite element stiffness matrix presented follows the formula-
tion found in C OOK , M ALKUS and P LESHA [1989]. The finite element stiffness matrix
Ke is given by

Ke = BT DB dV , (18.2)
Ve
where D is the stress–strain matrix and B is the strain–displacement matrix. The stiff-
ness matrix incorporates the material and geometrical stiffness of the element.
18.1.3. Stress–strain matrix

Stress is related to strain through the following relationship, for a linear material:
σ = Dε, (18.3)
 
εxx
 εyy 
 
ε 
where ε =  zz  (18.4)
 εxy 
 
εzy
εxz
and σ is a column vector with the respective strain components. For an isotropic mate-
rial, D is defined by
 υ υ 
1 1−υ 1−υ 0 0 0
 1 υ
0 0 0 
 1−υ 
 0 
Y (1 − υ)  1 0 0 
D=   ,(18.5)

(1 + υ)(1 − 2υ) 
1−2υ
0 0 
2(1−υ) 
 1−2υ 
 Symmetric 2(1−υ) 0 
1−2υ
2(1−υ)
where Y is the Young’s modulus, and υ is the Poisson ratio.
18.1.4. Strain–displacement matrix

The matrix B relates strains to displacements at the nodes
ε = Bqe , (18.6)
where ε is given above, and qe is the displacements at the element’s n nodes. For Carte-
sian coordinates, u, v and w are displacements in the x, y and z directions, respectively:
 
u1
 v1 
 
 w1 
 
 u2 
 
 v2 
qe =  
 w2  . (18.7)
 . 
 . 
 . 
 
 un 
v 
n
wn
The relation between strain and displacements at the nodes involve matrix multiplica-
tions as defined below. First, the definition of small strain can be written in matrix form
as
 
u,x
  u,y 
1 0 0 0 0 0 0 0 0   
0 0 0 0 1 0 0 0 0 u,z 
  v,x 

0 0 0 0 0 0 0 0 1 
ε=  
0 2 0 2 0 0 0 0 0 y 
1 1  v,

. (18.8)
   v, 
0 0 0 0 0 1 0 1 0 z 
2 2  w,x 
0 0 12 0 0 0 12 0 0  w, 
y
L: 6×9 w,z
The derivatives of displacement with respect to global coordinates in this equation can
be related to derivatives of displacements with respect to local coordinates by the 9 × 9
matrix M, in the following equation:
    
u,x 11 12 13 0 0 0 0 0 0 u,ξ
 u,y   21 22 23 0 0 0 0 0 0   u,η 
    
 z   31
u, 32 33 0 0 0 0 0 0   u,ζ 
    
 v,x   0 0 0 11 12 13 0 0 0   v,ξ 
    
 v,y  =  0 0 0 21 22 23 0 0 0   v,η  .
    
 v,z   0 0 0 31 32 33 0 0 0   v,ζ 
    
 w,x   0 0 0 0 0 0 11 12 13   w,ξ 
    
w,y 0 0 0 0 0 0 21 22 23 w,η
w,z 0 0 0 0 0 0 31 32 33 w,ζ

M: 9×9
(18.9)
The matrix M is composed of a 3 × 3 matrix, , defined later. The derivatives of dis-

placements with respect to local coordinates can now be related to the nodal displace-
ments (vector qe ) via the shape functions. We know that the interpolation functions
define the displacement in the element in terms of nodal displacements. For example,
the displacement in the x-direction, u, is

n
u= Ni ui = N1 u1 + N2 u2 + · · · + Nn un (18.10)
i=1
and the derivative with respect to a local coordinate, ξ , is
∂u n
= u,ξ = Ni,ξ ui = N1,ξ u1 + N2,ξ u2 + · · · + Nn,ξ un . (18.11)
∂ξ
i=1
This relation follows for all three displacement derivatives taken with respect to the
three local coordinates. It is written in matrix form as
   
u,ξ N1,ξ 0 0 N2,ξ 0 0 ... Nn,ξ 0 0
 η   1,η
u, N 0 0 N 2,η 0 0 ... Nn,η 0 0 
   
 ζ   1,ζ
u, N 0 0 N 2,ζ 0 0 ... Nn,ξ 0 0 
   
 ξ  
v, 0 N 1,ξ 0 0 N 2,ξ 0 ... 0 Nn,ξ 0 
   
 v,η  =  0 N1,η 0 0 N2,η 0 ... 0 Nn,η 0  qe .
   
 v,ζ   0 N1,ζ 0 0 N2,ζ 0 ... 0 Nn,ξ 0 
   
 w,ξ   0 0 N1,ξ 0 0 N2,ξ . . . 0 0 Nn,ξ 
   
w,η 0 0 N1,η 0 0 N2,η . . . 0 0 Nn,η
w,ζ 0 0 N1,ζ 0 0 N ... 0 0 Nn,ξ
2,ζ
Q: 9×(n∗ 3)
(18.12)
Thus, there are three matrix multiplications and the matrix B is composed of three
matrices. Comparing the equations, we get
B =
L
M Q . (18.13)

6×(n∗ 3) 6×9 9×9 9×(n∗ 3)
18.1.5. Derivation of inverse Jacobian

The Jacobian relates derivatives in the global coordinate system (x, y, z) to derivatives
in the local coordinate system (ξ, η, ζ ) and can be simply derived using the chain-rule.
So for a function Ψ (x, y, z):

ψ,ξ x,ξ y,ξ z,ξ ψ,x
ψ,η = x,η y,η z,η ψ,y . (18.14)
ψ,ζ x,ζ y,ζ z,ζ ψ,z

J: 3×3
In the isoparametric formulation, a point (x, y, z) in the element is interpolated from
the nodes with the shape functions:

n
n
n
x= Ni xi , y= Ni yi , z= Ni zi , (18.15)
i=1 i=1 i=1
where the shape functions are functions of the local coordinates. Therefore, the elements
in the Jacobian matrix can be evaluated by taking the derivative of the appropriate term
in the last equation with respect to the appropriate local coordinates:
 n n n 
i=1 Ni,ξ xi i=1 Ni,ξ yi Ni,ξ zi
n n i=1
J =  i=1 Ni,η xi Ni,η yi n
Ni,η zi  . (18.16)
n i=1
n i=1
n
N
i=1 i,ζ ix N
i=1 i,ζ i y N
i=1 i,ζ iz
This formula for J can be expanded as follows:
 
x1 y1 z1
 x2 y2 z2 
N1,ξ N2,ξ N3,ξ N4,ξ . . . Nn,ξ   x3 y3 z3 

J = N1,η N2,η N3,η N4,η . . . Nn,η  x4 y4 z4  
. (18.17)
N1,ζ N2,ζ N3,ζ N4,ζ . . . Nn,ζ  . . . 
 .. .. .. 
Dn : 3×n
xn yn zn

x: n×3
The matrix x contains the deformed positions of the n nodes in the element. The matrix
is finally given by

11 12 13
= J−1 = 21 22 23 . (18.18)
31 32 33
18.2. Modeling linear triangle isoparametric elements
Fig. 18.1 shows a triangular element P1 P2 P3 in a local orthonormal coordinate system

(α, β) with basis unit vectors α̂, β̂ and its origin at node P1 .
The local node displacement vectors are given by (pi , qi ) for every node i. The
real-world 3D-coordinate system is given by (x, y, z) and the real-world 3D node dis-
placements are given by (u, v, w). The displacement field is given by p = p(α, β) and
F IG . 18.1. Linear 3-node triangle element.

q = q(α, β). They are each interpolated from the nodal displacement degrees of free-
dom pi and qj :

p ξ 0 ξ2 0 ξ3 0
= Nqe , where N = 1 ,
q 0 ξ1 0 ξ2 0 ξ3
 
p1
 q1 
 
p 
qe =  2  . (18.19)
 q2 
 
p3
q3
ξi are the local area coordinates of any point P inside the triangle element: P divides
the triangle element into three sub-areas A1 , A2 and A3 . Area coordinates are defined
as ratios of areas:
A1 A2 A3
ξ1 = , ξ2 = , ξ3 = , (18.20)
A A A
where A is the area of the triangle element. Now since A = A1 + A2 + A3 , the ξi are
not independent, and they satisfy the constraint equation,
ξ1 + ξ2 + ξ3 = 1. (18.21)
This constraint equation and the linear relation between Cartesian (α, β) and area
(ξ1 , ξ2 , ξ3 ) coordinates are expressed by the following equations:

1 ξ1 ξ1 1
−1
α = A ξ2 , ξ2 = A α , (18.22)
β ξ3 ξ3 β
where

1 1 1 1 (α2 β3 − α3 β2 ) β23 α32
−1
A = α1 α2 α3 , A = (α3 β1 − α1 β3 ) β31 α13 (18.23)
2A
β1 β2 β3 (α1 β2 − α2 β1 ) β12 α21
with αij ≡ αi − αj and βij ≡ βi − βj , and 2A = det(A) = α21 β31 − α31 β21 .
Since we know that strains are defined as ε = Bqe , and using the chain rule, we obtain
the following expression for matrix B, after simplification:

1 β23 0 β31 0 β12 0
B = ∂N = 0 α32 0 α13 0 α21 . (18.24)
2A α β α β α β
32 23 13 31 21 12
18.2.1. Material properties matrix E

E is defined as

Y (1 − υ) υ 0
E= υ (1 − υ) 0 , (18.25)
(1 + υ)(1 − 2υ) 0 0 (1−2υ)
2
where Y = Young’s modulus of elasticity, υ = Poisson ratio.
18.2.2. Element stiffness matrix formulation

Finally, the element stiffness matrix Ke is given by

Ke = BT DB dV . (18.26)
Ve
But B and D are constant over the triangle element, and we suppose that the element
thickness t (which corresponds to the skin’s thickness) is also constant. Then Ke can be
simply written as
Ke = At ·
BT B ,
D (18.27)

6×6 6×3 6×6 3×6
where A is the area of the triangle element.
18.2.3. Forces on element nodes

The forces generated on the element nodes P1 , P2 and P3 are then written as
   
αF 1 p1
 βF 1   q1 
   
α  p 
Fe = Ke qe ⇐⇒  F 2  = Ke  2  . (18.28)
 βF 2   q2 
   
αF 3 p3
βF 3 q3
18.2.4. Expression for Ke in the real world coordinates

The node forces generated (αF i , βF i ), the node displacement vectors (pi , qi ) and the
node coordinates (αi , βi ) need to be transformed from and to the real-world 3-dimen-
sional coordinate system (x, y, z). The basis vectors α̂(xα̂ ; yα̂ ; zα̂ ) and β̂(xβ̂ ; yβ̂ ; zβ̂ ) in
the (x, y, z) system are given by the following:
→
−α −−→
α̂ = − → , where − →α = P1 P2 ,
α
→
− −−→ −−→
β → −−→ (P1 P2 · P1 P3 ) −−→
−
β̂ = − → , where β = P 1 P3 − −−→ P1 P2 . (18.29)
β P1 P2 2
18.2.5. Coordinates of element nodes

The coordinates of the element nodes are given as follows:
 −−→ −−→

 α = |P1 P2 · P1 P3 | ,
α1 = 0, →
−
α2 = α , 3 −−→
P1
β1 = 0,
P2
β2 = 0,
P3
 P1 P2 (18.30)
 →
−
β3 = β .
18.2.6. Coordinates of the nodal displacement field

The coordinates of the nodal displacement field are given as follows:

ui ui
pi = vi [ xα̂ yα̂ zα̂ ] , qi = vi [ xβ̂ yβ̂ zβ̂ ] . (18.31)
wi wi
F IG . 18.2. Triangular pyramid.
18.2.7. Real-world coordinates of the resulting nodal forces

The real-world coordinates of the resulting nodal forces are given by

xF i xα̂ xβ̂
αF i
yF i = yα̂ yβ̂ . (18.32)
βF i
zF i zα̂ zβ̂
18.3. Element volume calculations
Calculating the volume of a hexahedral element is used to verify the incompressibil-

ity properties of the model, depending on the value of Poisson’s ratio: we calculate the
volume of every element after every displacement iteration and we verify volume con-
servation. Volume calculations are also necessary when including body forces in the
model, such as gravity.
18.3.1. Volume of a triangular pyramid

In an orthonormal basis, centered at C, the volume of the pyramid CGEF is given by
(see Fig. 18.2)
1 −→ −→ −→
VEFG = CE · (GE × GF ) . (18.33)
6
Analytically, we get

(Ex − Cx ) · (Ey − Gy ) · (Fz − Gz ) − (Ez − Gz ) · (Fy − Gy )
1

VEFG = + (Ey − Cy ) · (Ez − Gz ) · (Fx − Gx ) − (Ex − Gx ) · (Fz − Gz ) .
6

+ (Ez − Cz ) · (Ex − Gx ) · (Fy − Gy ) − (Ey − Gy ) · (Fx − Gx )
(18.34)
18.3.2. Volume of a hexahedral element
First, we find the center point of the element (see Fig. 18.3):
1
Cx = (P1x + P2x + P3x + P4x + P5x + P6x + P7x + P8x ),
8
1
Cy = (P1y + P3y + P5y + P7y ),
4
1
Cz = (P1z + P5z ), (18.35)
2
where Pi are the nodes of the element.
F IG . 18.3. Volume calculation of a hexahedral element.
F IG . 18.4. Volume calculation of a hexahedral element using only 5 triangular pyramids.
Finally, the total volume of the element is calculated by dividing the element into
several different triangular pyramids, and calculating the volume of each one of them:
VE = (V451 + V485 + V234 + V124 + V587 + V687 + V326 + V673 + V784

+ V437 + V156 + V162). (18.36)
We can actually express the volume of a hexahedral element using only 5 triangular
pyramids, by dividing the element as shown in Fig. 18.4. Then the total volume of the
element is given by
VE = (VEFGD + VEFGH + VCEGH + VBFHE + VHAGF ). (18.37)
18.4. Is a data point inside or outside an element?
The ability to determine whether a point is inside or outside of an element is very im-
portant especially when determining the tissue type of an element: all the data points
F IG . 18.5. Is C inside the hexahedron?
from the segmented breast image which lie inside the hexahedral element have to be
counted, in order to determine what is the highest percentage of tissue type inside the
element.
In 2 dimensions.
If PA × PD, PD × PC,
PC × PB, PB × PA,
all have the same direction,
then P is inside element ABCD;
otherwise P is outside ABCD.
In 3 dimensions. C is inside the hexahedron, if and only if all of the following quan-
tities are negative (see Fig. 18.5):
−−→ −−→ −−→ −−→ −−→ −−→ −−→ −−→
N214 · P2 C; N567 · P5 C; N432 · P4 C; N785 · P7 C
−−→ −−→ −−→ −−→ −−→ −−→ −−→ −−→
N158 · P1 C; N237 · P2 C; N841 · P8 C; N762 · P7 C
−−→ −−→ −−→ −−→ −−→ −−→ −−→ −−→
N348 · P3 C; N126 · P1 C; N873 · P8 C; N651 · P6 C, (18.38)
−−→
where Pi are the nodes of the element, and Nij k is the outer unit normal of triangle
Pi Pj Pk .
18.5. Silicon phantom construction
The geometry of the deformable phantom consists of a rectangular box (84 × 82 ×

70 mm) containing a rectangular inclusion (20 × 23 × 20 mm), which is 4.3 times stiffer
than the surrounding silicon (Fig. 10.1). The gel system is composed of two parts, cata-
lyst (part A) and resin (part B), the ratio of which determines its elastic properties (a de-
crease in A:B produces stiffer gels). The components, the catalyst (part A) and the resin
(part B), both contain silicon copolymers which form cross-links when combined. Since
only approximately 2% of the material becomes cross-linked, the material becomes a
gel, and its stiffness is directly proportional to the amount of cross-linking (G OLD -
STEIN , K UNDEL , DAUBE -W HITERSPOON , T HIBAULT and G OLDSTEIN [1987]). Parts
A and B of the gel system were mixed in a ratio 1:1.7. This provided a good combi-
nation of flexibility under compression and integrity under gravity, for the surrounding
silicon in the phantom. Parts A and B of the gel system were then mixed in a ratio 1:5.7,
for the stiff inclusion. The phantom was built as follows:
1. The molds consist of two rectangular boxes, one of size (84 × 82 × 70 mm) to
house the whole silicon phantom, and another smaller mold of size (20 × 23 ×
20 mm) used to make the stiff silicon inclusion. Both molds are made of heat-
resistant PVC. The inside walls of the molds were sprayed with Pam oil, and cov-
ered with sheets of transparent plastic wrapper, making it much easier to remove
the silicon phantoms out of the molds when ready.
2. Parts A and B of the gel system were mixed in a ratio 1:1.7. This provided a good
combination of flexibility under compression and integrity under gravity. The mix
was stirred for 5 minutes, and then poured into the larger mold so as to fill about
half of the mold.
3. Parts A and B of the gel system were then mixed in a ratio 1:5.7, stirred for 5
minutes and then poured into the smaller mold in order to fill it up.
4. Both molds were then heated up at 175 ◦ F for 36 hours, the time needed for the
gel systems to cure. The heat acted as a catalyst in the curing process.
5. The stiff silicon gel inclusion was removed from its mold, and then placed inside
the larger mold on the top of the already cured silicon gel.
6. Parts A and B of the gel system were again mixed in a ratio 1:1.7, stirred for 5
minutes and then carefully poured into the larger mold in order to fill it up.
7. The large mold was heated up again at 175 ◦ F during 36 hours.
8. The full silicon gel phantom was finally removed from its mold by pulling on the
plastic wrap paper, and secured inside a custom-built pressure device (also made
of PVC material, which does not cause any extraneous signal when imaged in the
MR machine).
Acknowledgements
The authors are thankful to Norm Butler, Allen Bonner, Idith Haber, Reid Miller, Bruno
Carvalho and Joe Giammarco for their help in various aspects of this work.
References
ABAQUS/Standard V.5.8 (1998). Hibbitt, Karlsson & Sorensen, vol. II: 14.1.4–1, 14.1.4–17.
A GACHE , P.G., M ONNEUR , C., L EVEQUE , J.L., D E R IGAL , J. (1980). Mechanical properties and Young’s
modulus of human skin in vivo. Arch. Dermatol. Res. 269, 221–232.
A ZAR , F.S., M ETAXAS , D.N., M ILLER , R.T., S CHNALL , M.D. (2000). Methods for predicting mechanical
deformations in the breast during clinical breast biopsy. In: 26th IEEE Annual Northeast Bioengineering
Conference.
A ZAR , F.S., M ETAXAS , D., S CHNALL , M.D. (1999). A finite element model of the breast for predicting
mechanical deformations during interventional procedures. In: Proc. 7th Intl. Soc. Magn. Reson. Med.,
p. 1084.
A ZAR , F.S., M ETAXAS , D.N., S CHNALL , M.D. (2001). A deformable finite element model of the breast for
predicting mechanical deformations under external perturbations. J. Academic Radiology.
A ZAR , F.S., M ETAXAS , D.N., S CHNALL , M.D. (2002). Methods for modeling and predicting mechanical
deformations of the breast under external perturbations. Medical Image Analysis 6 (1), 1–27.
BAUMANN , R., G LAUSER , D. (1996). Force feedback for virtual reality based minimally invasive surgery
simulator. In: Medicine Meets Virtual Reality (IOS Press, Amsterdam).
B EHRENBRUCH , C.P., M ARIAS , K., A RMITAGE , P., YAM , M., M OORE , N., E NGLISH , R.E., B RADY , M.
(2000). MRI-mammography 2D/3D data fusion for breast pathology assessment. In: Proc. Medical Image
Computing and Computer Assisted Intervention, MICCAI, pp. 307–316.
B OETES , C., BARENTSZ , J.O., et al. (1994). MR characterization of suspicious breast lesions with a
gadolinium-enhanced turboFLASH subtraction technique. Radiology 193, 777–781.
C ARVALHO , B.M., G AU , C.J., H ERMAN , C.T., K ONG , T.Y. (1999). Algorithms for fuzzy segmentation.
Pattern Anal. Appl. 2, 73–81.
C HADWICK , J., H AUMANN , D., PARENT, R. (1989). Layered construction of deformable animated charac-
ters. Computer Graphics (SIGGRAPH’89) 23, 243–252.
C HEN , D.T., Z ELTZER , D. (1992). Pump it up: Computer animation of a biomechanically based model of
the muscle using the finite element method. Computer Graphics (SIGGRAPH’92) 26, 89–98.
C OOK , R.D., M ALKUS , D.S., P LESHA , M.E. (1989). Concepts and Applications of Finite Elements Analysis
(Wiley, New York).
C OTIN , S., D ELINGUETTE , H., AYACHE , N. (1999). Real-time elastic deformations of soft tissues for surgery
simulation. IEEE Trans. Visualization and Computer Graphics 5 (1), 62–73.
C RISP , J.D.C. (1972). Properties of tendon and skin. In: Fung, Y.C. (ed.), Biomechanics: Its Foundations
and Objectives (Prentice Hall, New York).
D ERSHAW (1996). Stereotaxic breast biopsy. Seminars in Ultrasound, CT, and MRI 17 (5), 444–459.
D ESBRUN , M., G ASCUEL , M.P. (1995). Animating soft substances with implicit surfaces. Computer Graph-
ics (SIGGRAPH’95), 287–290.
D E V RIES , P.L. (1994). A First Course in Computational Physics (Wiley, New York), pp. 207–225.
D RONKERS (1992). Stereotaxic core biopsy of breast lesions. Radiology 183, 631–634.
E GAN , R.L. (1988a). Breast embryology, anatomy and physiology. In: Breast Imaging: Diagnosis and Mor-
phology of Breast Diseases, pp. 30–58, Chapter 4.
E GAN , R.L. (1988b). Malignant breast lesions. In: Breast Imaging: Diagnosis and Morphology of Breast
Diseases, pp. 227–231, Chapter 14.
E LDEN , H.R. (1977). Biophysical Properties of Skin (Wiley–Interscience, New York).
653
E RGATOUDIS , I., I RONS , B.M., Z IENKIEWICZ , O.C. (1968). Curved isoparametric, ‘quadrilateral’ elements
for finite element analysis. Int. J. Solids Structures 4 (1), 31–42.
F ERLAY , J., B RAY , F., et al. (2001). Globocan 2000: Cancer Incidence, Mortality and Prevalence Worldwide,
Version 1.0 (IARCPress, Lyon).
F ISCHER , U., V OSSHENRICH , R., D OLER , W., H AMADEH , A., O ESTMANN , J.W., G RABBE , E. (1995).
MR Imaging-guided breast intervention: experience with two systems. Radiology 195, 533–538.
F ISCHER , U., V OSSHENRICH , R., K EATING , D., B RUHN , H., D OLER , W., O ESTMANN , J.W., G RABBE ,
E. (1994). MR-guided biopsy of suspect breast lesions with a simple stereotaxic add-on device for surface
coils. Radiology 192, 272–273.
F UNG , Y.C. (1972). Stress–strain history relations of soft tissues in simple elongation. In: Fung, Y.C., Per-
rone, N., Anliker, M. (eds.), Biomechanics: Its Foundations and Objectives (Prentice Hall, Englewood
Cliffs, NJ).
F UNG , Y.C. (1981). Biomechanics: Mechanical Properties of Living Tissues (Springer-Verlag, New York),
pp. 203–212.
F UNG , Y.C. (1987). Mechanics of soft tissues. In: Skalak, R., Chien, S. (eds.), Handbook of Bioengineering
(McGraw–Hill, New York).
F UNG , Y.C. (1993). Biomechanics: Mechanical Properties of Living Tissues, second ed. (Springer-Verlag,
New York).
F UNG , Y.C. (1994). A First Course in Continuum Mechanics (Prentice Hall, Englewood Cliffs, NJ).
G ILLES , R., G UINEBRETIERE , J.M., et al. (1994). Non-palpable breast tumors: Diagnosis with contrast-
enhanced subtraction dynamic MR imaging. Radiology 191, 625–631.
G OLDSTEIN , D.C., K UNDEL , H.L., D AUBE -W HITERSPOON , M.E., T HIBAULT , L.E., G OLDSTEIN , E.J.
(1987). A silicone gel phantom suitable for multimodality imaging. Invest. Radiol. 22, 153–157.
G REEN , A.E., Z ERNA , W. (1968). Theoretical Elasticity (Oxford Univ. Press, London), p. 99.
H ABER , I., M ETAXAS , D., A XEL , L. (2000). Three-dimensional motion reconstruction and analysis of the
right ventricle using tagged MRI. Medical Image Analysis 4 (4), 335–355.
H ARMS , S.E., F LAMIG , D.P. (1994). Staging of breast cancer with MR imaging. Magn. Reson. Imaging
Clin. N. Am. 2, 573–584.
H ARMS , S.E., F LAMIG , D.P., et al. (1993). MR imaging of the breast with rotating delivery of excitation off
resonance: Clinical experience with pathologic correlation. Radiology 187, 493–501.
H ARRIS , J.R., L IPPMAN , M.E., M ORROW , M., H ELLMAN , S. (1996). Diseases of the Breast (Lippincott–
Raven).
H AYES , W.C., K EER , L.M., H ERMANN , G., M OCKROS , L.F. (1972). A mathematical analysis for indenta-
tion tests of articular cartilage. J. Biomechanics 5, 541–551.
H OU , F.J., L ANG , S.M., et al. (1998). Human vertebral body apparent and hard tissue stiffness. J. Biome-
chanics 31, 1009–1015.
J OUKHADAR , A. (1995). Energy based adaptive time step and inertia-matrix based adaptive discretization for
fast converging dynamic simulation. In: Proc. Intl. Workshop on Visualisation and Mathematics (Springer-
Verlag, Heidelberg).
K ABEL , J., VAN R IETBERGEN , B., et al. (1999). The role of an effective isotropic tissue modulus in the
elastic properties of cancellous bone. J. Biomechanics 32, 673–680.
K OJIC , M., BATHE , K.J. (1987). Studies of finite element procedures – stress solution of a closed elastic
strain path with stretching and shearing using the updated Lagrangian Jaumann formulation. Computers &
Structures 26 (1/2), 175–179.
K ROUSKOP , T.A., W HEELER , T.M., K ALLEL , F., G ARRA , B.S., H ALL , T. (1998). The elastic moduli of
breast and prostate tissues under compression. Ultrasonic Imaging 20, 151–159.
K UEHNAPFEL , U.G., N EISIUS , B. (1993). CAD-based graphical computer simulation in endoscopic surgery.
End. Surg. 1, 181–184.
L AWRENCE , A.J., ROSSMAN , P.J., M AHOWALD , J.L., M ANDUCA , A., H ARTMANN , L.C., E HMAN , R.L.
(1999). Assessment of breast cancer by magnetic resonance elastography. In: Proc. 7th Intl. Soc. Magn.
Reson. Med., p. 525.
L EE , G.C., T SENG , N.T. (1982). Finite element analysis in soft tissue mechanics. In: Gallagher, R.H., Simon,
B.R., Johnson, P.C., Gross, J.F. (eds.), Finite Elements in Biomechanics (Wiley, Chichester, UK).
References 655
L UCIANI , A., J IMENEZ , S., F LORENS , J.L., C ADOZ , C., R AOULT , O. (1991). Computational physics:
a modeler simulator for animated physical objects. In: Eurographics Workshop on Animation and Sim-
ulation, pp. 425–437.
M AUREL , W., W U , Y., M AGNENAT T HALMANN , N., T HALMANN , D. (1998). Biomechanical Models for
Soft Tissue Simulation. In: Basis Research Series, Esprit (Springer-Verlag, Berlin).
M ESEURE , P., C HAILLOU , C. (1997). Deformable body simulation with adaptive subdivision and cuttings.
In: Proc. Fifth Int. Conf. in Central Europe on Computer Graphics and Visualization (Pergamon Press,
Oxford), pp. 361–370.
M ETAXAS , D. (1992). Physics-based modeling of nonrigid objects for vision and graphics. PhD thesis, De-
partment of Computer Science, University of Toronto.
M ETAXAS , D., T ERZOPOULOS , D. (1993). Shape and nonrigid motion estimation through physics-based
synthesis. IEEE Trans. Pattern Analysis and Machine Intelligence 15 (6), 569–579.
M ILLER , G. (1988). The motion dynamics of snake and worms. Computer Graphics (SIGGRAPH’88) 23,
169–173.
NCI (1998). Understanding Breast Cancer Treatment (National Cancer Institute), pp. 6–7, NIH 98-4251.
N ORTON , A., T URK , G., BACON , B., G ERTH , J., S WEENEY , P. (1991). Animation of fracture by physical
modeling. The Visual Computer 7, 210–219.
O REL , S.G., S CHNALL , M.D., N EWMAN , R.W., P OWELL , C.M., T OROSIAN , M.H., ROSATO , E.F. (1994).
MR imaging-guided localization and biopsy of breast lesions: Initial experience. Radiology 193, 97–102.
PARK , J., M ETAXAS , D., Y OUNG , A.A., A XEL , L. (1996a). Analysis of left ventricular wall motion based
on volumetric deformable models and MRI-SPAMM. Medical Image Analysis 1 (1), 53–71.
PARK , J., M ETAXAS , D., Y OUNG , A.A., A XEL , L. (1996b). Deformable models with parameter functions
for cardiac motion analysis from tagged MRI data. IEEE Trans. Medical Image Processing.
PARKER , S.H., L OVIN , J.D., J OBE , W.E., et al. (1990). Stereotactic breast biopsy with a biopsy gun. Radi-
ology 176, 741–747.
P ICINBONO , G., D ELINGUETTE , H., AYACHE , N. (2001). Non-linear and anisotropic elastic soft tissue
models for medical simulation. In: IEEE Int. Conf. on Robotics and Automation, ICRA2001, Seoul, Korea.
P RESS , W.H., T EUKOLSKY , S.A., V ETTERLING , W.T., F LANNERY , B.P. (1992). Numerical Recipes in C:
The Art of Scientific Computing (Cambridge Univ. Press, Cambridge), pp. 707–725.
R EDDY , N.P., S ONG , G.J. (1995). Tissue cutting in virtual environments. In: Medicine Meets Virtual Reality
IV (IOS Press), pp. 359–364.
S AHA , P.K., U DUPA , J.K., O DHNER , D. (2000). Scale-based fuzzy connected image segmentation: theory
algorithms, and validation. Computer Vision and Image Understanding 77, 145–174.
S ARVAZYAN , A.P., S KOVORODA , A.R., E MELIANOV , S.Y., F OWLKES , J.B., P IPI , J.G., A DLER , R.S.,
B UXTON , R.B., C ARSON , P.L. (1995). Biophysical Bases of Elasticity Imaging (Plenum Press, New
York).
S CHNEIDER , D.C., D AVIDSON , T.M., N AHUM , A.M. (1984). In vitro biaxial stress–strain response of hu-
man skin. Arch. Otolaryngol. 110, 329–333.
S CIARETTA , J., B ISHOP , J., S AMANI , A., P LEWES , D.B. (1999). MR validation of soft tissue deformation
as modeled by non-linear finite element analysis. In: Proc. 7th Intl. Magn. Reson. Med., p. 246.
S KOVORODA , A.R., K LISHKO , A.N., G USAKYAN , D.A., M AYEVSKII , Y.I., Y ERMILOVA , V.D., O RAN -
SKAYA , G.A., S ARVAZYAN , A.P. (1995). Quantitative analysis of the mechanical characteristics of patho-
logically changed soft biological tissues. Biophysics 40 (6), 1359–1364.
S PEETER , T.H. (1992). Three-dimensional finite element analysis of elastic continua for tactile sensing. Intl.
J. Robotics Res. 11 (1), 1–19.
S PENCER , A.J.M. (1980). Continuum Mechanics (Longman, London), pp. 153–163.
S TAVROS , A.T., T HICKMAN , D., R APP , C.L., D ENNIS , M.A., PARKER , S.H., S ISNEY , G.A. (1995). Solid
breast modules: use of sonography to distinguish between benign and malignant lesions. Radiology 196,
123–134.
S ZEKELY , G., B RECHBUHLER , C H ., H UTTER , R., R HOMBERG , A., S CHMID , P. (1998). Modelling of soft
tissue deformation for laparoscopic surgery simulation. In: Medical Image Computing and Computer-
Assisted Intervention, MICCAI, pp. 550–561.
T ERZOPOULOS , D., P LATT , J., et al. (1987). Elastically deformable models. Computer Graphics (SIG-
GRAPH’87) 21, 205–214.
T OFTS , P.S., B RIX , G., et al. (1999). Estimating kinetic parameters from dynamic contrast-enhanced T(1)-
weighted MRI of a diffusible tracer: standardized quantities and symbols. J. Magn. Reson. Imaging 10,
223–232.
V ERONDA , D.R., W ESTMANN , R.A. (1970). Mechanical characterization of skin-finite deformations. J. Bio-
mech. 3, 111–124.
W ELLMAN, P.S. (1999). Tactile imaging. Thesis. Harvard University, Cambridge, MA.
W ELLMAN , P.S., H OWE, R.D. (1998). Harvard Bio-Robotics Lab. Tech. Report. #98-121.
W ILLIAMS , C., C LYMER , B., S CHMALBROCK , P. (1999). Biomechanics of breast tissue: preliminary study
of force-deformation relationship. In: Proc. 7th Intl. Soc. Magn. Reson. Med., p. 524.
YAMADA , H. (1970). Strength of Biological Materials (Williams & Wilkins, Baltimore, MD).
Y OUNG , A.A., A XEL , L., D OUGHERTY , L., B OGEN , D.K., PARENTEAU , C.S. (1993). Validation of tag-
ging with MR imaging to estimate material deformation. Radiology 188, 101–108.
Y UCESOY , C.A., K OOPMAN , B., et al. (2002). Three-dimensional finite element modeling of skeletal muscle
using a two-domain approach: linked fiber-matrix mesh model. J. Biomechanics 35, 1253–1262.
Z HANG , M., Z HENG , Y.P., M AK , A.F. (1997). Estimating the effective Young’s modulus of soft tissues from
indentation tests – nonlinear finite element analysis of effects of friction and large deformation. Med. Eng.
Phys. 19 (6), 512–517.
Z HUANG , Y., C ANNY , J. (1999). Real-time and physically realistic simulation of global deformation. In:
SIGGRAPH’99.
Z IENKIEWICZ , O.C. (1977). The Finite Element Method, third ed. (McGraw–Hill, London).
Z IENKIEWICZ , O.C., TAYLOR , R.L. (1989). The Finite Element Method, fourth ed. (McGraw–Hill, New
York).
Subject Index
3-node triangular isoparametric elements, 601 – time-varying elastance, 142

3D mesh, 597, 600, 602, 603 – Tozeren, 142
3D visualization, 601 – Wong, 142
active transport, 132
ABAQUS, 597, 599, 601, 608, 612, 614, 641 adaptive meshing, 172, 173
abdomen adaptive step sizing, 604
– anatomy, 324 ADP, 146
– bar impact, 329 advection–diffusion problem, 51
– cavity, 295, 324 airbag models
– organs, 410, 430 – contact, 371
– validation, 295 – gas model, 421
absolute refractory period, 139 – inflow, 421
acceleration, 22 – leakage, 421
accelerometers, 256, 260, 300, 318 – vents, 421
accident reconstructions, 240 airbags
accidentology, 296 – driver airbag simulation, 281
accuracy, 596, 599, 606, 608, 609, 638, 641 – female driver OOP simulation, 361
acetabulum, 343–345 – OOP simulation (6 year old), 281
Achilles tendon, 347, 350, 405 – passenger OOP simulation, 304
ACL, 386 AIS injury scale, 243
actin, 142–144 Arbitrary Lagrangian Eulerian (ALE)
action potential, 130–132, 136, 138, 139 formulation, 71
– duration, 132, 133, 146, 147 anastomosis, 113
– heterogeneity, 133 anatomical models
– morphology, 147 – canine, 163, 168
– propagation, 133 – fiber architecture, 167
active fiber stress development, 142 – porcine, 165
active force development, 139, 142 – rabbit, 164, 165, 167
active muscle action anatomy
– energy, 276 – abdomen, 324
– Hill model, 413 – foot/ankle, 349
– Hill model bracing, 276 – lower extremity, 342, 343
– mechanical properties, 275, 400, 417 – neck, 311
– neck, 314–316 – pelvis, 342
active tension models, 142 – upper extremity, 335, 336
– fully history-dependent, 142 angiogenesis, 189–191, 194
– Hill, 142 animal test, 241
– Panerai, 142 anisotropic propagation, 145, 146
657
658 Subject Index
– orthotropic, 146 – inter-vertebral discs, 399

– transversely isotropic, 146 – inter-vertebral joints, 395, 398
anisotropy, 638 – internal organs, 404
anterior cruciate ligament (ACL), 343 – joints (synovial), 389
aortic valve, 163 – muscles, 400
arachnoid membrane, 306 – skin, 404
arm–shoulder complex, 274 – tendons, 405
– anatomy, 335 biomechanical model, 551
– model, 335 BioRID, 257, 258
– muscle activation, 278 BioSID, 254, 256, 257
arterioles, 10 bladder, 242, 246, 324, 326, 412
ASDH (see brain injury), 260 bone
assembly, 155, 170, 172, 173 – cortical, 250
Athens University, 297 – – dynamic impacts, 355
atlanto-axial joint, 315 – – ribs, 293
atlas (C1), 395 – – shell mesh on solids, 351
ATP, 132 – – static crush behaviour, 354, 355
ATPase, 141 – metatarsal, 350
atrioventricular node, 145 – pelvic bone, 292, 343, 345
atrioventricular valves, 163 – pubis, os pubis, 342
automaticity, 144, 145 – trabecular, 250
avascular, 189–191, 193 – – dynamic impacts, 355
axis (C2), 395
– – skull (diploe), 303
– – solid mesh in shells, 351
bag-slap, 361
– – static crush behaviour, 354, 355
bandwidth, 364
bone materials
basis functions, 152, 156, 157, 176, see also:
– age dependency of properties, 380
interpolation functions
– cortical
– cubic Hermite, 176, 177
– – orthotropic stiffness matrix, 375, 377
– linear Lagrange, 175
– one-dimensional, 178, 180 – – transverse isotropic stiffness, 377
– quadratic Lagrange, 175, 176 – elastic moduli and ultimate stress, 376
– three-dimensional, 180 – quasi-static, 375
– two-dimensional, 177, 178 – strain rate dependent properties, 378
Bayesian Estimation, 559 – trabecular
belt loading, 329, 333 – – compressive properties, 378
biceps, 275, 278, 335 – – orthotropic stiffness, 378
bidomain field equations, 151, 154, 169, 171, – – shear properties, 380
172 boundary condition
bidomain model, 149–151, 173 – Dirichlet, 32, 38
bilinear, 198 – Neumann, 32, 38
bio-bag models boundary conditions, 130, 151–154, 156, 465,
– abdominal organs, 295 472, 477, 478, 483, 498–500, 502, 503,
– body cavity models, 430 507, 508, 511, 526, 543, 600, 601, 612,
– heart, 423, 424 616, 618, 622, 623, 627, 631, 634, 637,
– lungs, 422 641, 642
– mediastinum, 423, 424 – essential, 154
– mediastinum and heart, 423, 424 – natural, 155, 169, 170
– thoracic organs, 293 boundary element modeling, 476, 506, 506
biological tissues, 606 boundary value problem, 154
biomaterials, 241, 245, 247–251, 375, 412 brachialis, 335
– articular cartilage and meniscus, 399 bracing, 400
– bone, 248, 250, 375 brain
– brain, 389 – bridging veins, 260–262
Subject Index 659
brain injury calibration

– Acute Subdural Hematoma (ASDH), 260 – female thorax, 357
– concussion, 243, 262, 308, 309 – foot ligaments, 351
– contusion, 260 – H-Ankle&Foot, 352
– coup & countercoup, 308 – H-ARB, 299
– diffuse, 262, 309 – H-Head, 306, 307
– Diffuse Axonal (DAI), 260, 262, 309 – H-LE, 345
– hematoma – H-UE, 338
– – epidural, 243 calmodulin, 138
– – subdural (SDH), 243, 262, 310 calsequestrin, 138
– neural damage, 243 cancerous tissue, 596, 603
– SIMon criteria, 260 capacitance, 135, 148
brain materials capillaries, 10
– brain stem, 394 carcinomas, 620
– bridging veins, 394 cardiac conduction system, 145
– cerebellum, 394 cardiac cycle, 163
– CSF (cerebro-spinal fluid), 393 cardiac muscle testing, 139, 140
– dura mater, 394 cartilage
– general, 242, 389, 390 – general, 246, 247
– pia mater, 394 – lower extremity joints, 337, 343, 345, 346, 354
– tentorium, 394 – material properties, 399
– white and gray matter, 393 catalyst, 651
brain matter, 260, 305, 307 Cauchy stress, 27
brain models
Cauchy stress tensor, 28
– H-Head model, 302, 311
Cauchy’s infinitesimal strain tensor formulation,
– Madymo model, 267
605
– Radioss model, 269
cell-to-cell coupling, 133
– validation, 308
central difference scheme, 366
– WSU model, 262
cerebellum, 306, 307
brain shift, 568
cerebrospinal fluid (CSF), 307, 393
– compensation, 569
cervical spine
breast positioning, 635, 636, 642
– general, 243
breast tissue, 599, 606, 608, 616, 619–622, 638,
– H-Model, 311
642
breasts, 357, 358, 361, 404 – HUMOS model, 291
BreastView, 596, 597, 601, 602, 608, 611, 612, – modeling aspects, 313
614, 618, 631, 634, 641 CFD (computational fluid dynamics), 363
Brite–EuRam, 248, 296, 375 Chalmers University, 257, 297, 375
buffers, 138 characteristic curves, 97
bundle branches, 145 characteristic variables, 96, 103
bundle of His, 145 chest band, 329
chest belt loading, 329
C0 -continuity, 365, 640 chordae tendineae, 163
C1 -continuity, 365 circulatory system, 9
cable equation, 149 – large circulation, 8
cable theory, 147–149 – small circulation, 8
CAD, 283, 288, 289, 296, 299, 313 cisternae, 138
cadaver tests, 240, 241, 266, 283, 320, 328, 329, clavicle, 274, 292, 336, 375
341, 361, 375 clavicula, 274, 335
calcaneus, 350 cleavage planes, 168
calcium concentration, 136, 138–140, 142–144, coccyx, 323, 342
174 coercive, 216, 219
calcium induced calcium release, 138 collateral ligaments, 343
calcium transient, 142, 143 collocation, 152, 153
660 Subject Index
compact embedding, 201, 215 convergence analysis, 598, 625, 627, 628, 638
compatibility conditions, 100 convex, 216–218, 226
completeness, 639 cooperativity, 144
compliance, 10 Cooper’s ligaments, 596, 620, 622, 623, 638,
compliant vessel, 111 641
compound trapezium rule, 223 coordinate systems, 159–162
compression, 596–600, 603, 604, 606, 608, 609, – Cartesian, 160
612–618, 624–626, 629, 631, 632, 634, – curvilinear, 159
636–638, 641, 642, 652 – element, 156
computational fluid dynamics (CFD), 363 – fiber aligned, 162
computational structural mechanics (CSM), 363 – fiber-sheet, 169
concentration gradient, 132 – global, 156, 157
conductivity tensor, 147, 149, 150, 169, 170 – local, 157, 159
condyle split, 343, 345 – orthogonal curvilinear, 162
conformity, 640 – prolate spheroidal, 162, 164
connective tissue, 620 coordinate transformation, 158, 159, 162
connectivity matrix, 157, 170 coraco-brachialis, 335
conservation coronal plane and section, 271, 272
– of charge, 147–150 coronary by-pass, 15
– of current, 135, 149 coronary vessels, 163
– of mass, 131 correspondence estimation
– of momentum, 131 – robust point matching, 570
constant density fluid, 26 – shape-based, 580
constitutive law, 30 cortical bone characteristics, 353
contact CRABI, 254, 258
– algorithms, 238, 301, 346, 370, 372, 427 crash codes, 238
– events, 238, 370–372 cross-links, 651
– interface, 273, 274, 301, 337, 345, 354, 370 crossbridge, 130, 139, 141–144
– inter-vertebral contact, 313 – distortion, 144
– master segment, 372 – recruitment, 139, 144
– nodal constraints, 372 crossbridge force–velocity relation, 141
– options, 371, 372, 427, 428, 430, 431 cruciform ligaments, 315
– penalty contact, 372 crushable foam material, 353
– penetration, 265, 372, 373, 429 CSDM (Cumulative Strain Damage Measure),
– perforation, 373 260
– rigid walls, 372 CSF (cerebrospinal fluid), 306, 307, 393
– slave node, 372 CSM (computational structural mechanics), 363
– sliding interfaces, 275, 313, 429 cuboid, 349
– thickness, 370, 373 cyst, 598, 603, 615, 624, 625, 627, 628, 630,
– treatment, 370 634, 636–638, 640, 641
continuity, 157, 159, 176
continuity equation, 26 DAI (see brain injury), 260, 306
continuous dependence, 209 damageable bones, 337, 345, 353
continuous embedding, 205 damping matrix, 366
continuous form, 198 dash board, 343
continuous function, 195, 196, 204, 211, 216, dashpot, 367, 410, 413, 416
218 Data Extraction, 601
continuum mechanics, 553 data parallelism, 172
contractile activation, 139 data structure, 465, 502, 506, 507, 516–518,
contraction, 174 522, 535, 542
contrast-enhancing agent, 594 DDM (Dilatation Damage Measure), 260
controllability, 189, 190, 215 deformable breast model, 599
convective term, 29 deformable model, 596–598, 602, 608, 618, 634,
– numerical treatment, 45 642
Subject Index 661
deformation, left ventricular551, 552 elasticity, 600, 622, 648

– brain, 572 electrocardiogram, 146, 171
– cardiac, 576 electrochemical gradient, 132
– strain, 576 element parameterization, 158
deformation gradient, 22, 58, 592, 607, 610 ellipses of revolution, 163, 164
degrees of freedom, 603, 605, 616 end-systolic pressure, 140
deltoid, 335, 388 entropy flux, 106
dependent variable interpolation, 156 entropy function, 106
depolarization, 131–133, 138, 139, 145–147 epiphysis, 343
deviatoric strain invariants, 610 equations of motion, 238, 255, 363, 364, 366,
deviatoric stretches, 610 370, 421
diagnosis, 596, 597 equilibrium, 132, 139
diagonal mass matrix, 364 ESI Software, 254, 255, 257, 269, 271, 275, 280,
diastole, 12 293, 295, 297, 298, 363, 421–424
diffusion, 132, 192, 193, 195, 203 Euler expansion, 23
DIGIMATION, 271, 280, 281, 283, 299, 313 Euler integration technique, 605
Dirac delta function, 153 Euler method, 605
discretization, 222, 224 Eulerian variables, 19, 71
– errors, 596, 606, 608, 637, 639 European Commission, 248, 251, 296, 375
discs, 301, 313, 320, 321, 323, 325, 329, 334, EuroSID, 239, 253, 254, 256, 257, 265
399, 412 excitation–contraction coupling, 130, 174
– materials, 401, 402 existence, 194, 196, 198, 201, 204, 216
dislocations, 324, 325, 337, 343 explicit integration scheme, 367
displacement degrees of freedom, 591, 603, 604, explicit methods, 364
616, 647 explicit scheme, 222, 224
displacement difference, 625, 637
external forces, 591, 603, 604
displacement iteration, 596, 600, 605, 606, 612,
616, 638, 649
falx cerebri, 261
displacement method, 363
fat, 596, 601, 603, 619–622, 625, 630, 638, 641
displacement steps, 596, 606
– compartmentalization, 596, 641
displacement vector, 60
fatty tissue, 596, 597, 619, 622, 623, 638, 641
distal, 14
FEM, 597, 598, 601, 608, 612, 641
domain discretization, 155, 158, 170, 172
dorsiflexion, 269, 351, 355 female H-Model, 357
ductal carcinoma, 620 femoral condyle, 343, 345
dummies femur
– 5th percentile female, 357, 358 – anatomy, 343
– BioRID rear-impact dummy, 258 – femur neck, 343
– CRABI child restraint dummy, 258 – injury, 343, 345
– description, 253 – mechanical properties
– difference with humans, 301, 361 – – age dependent, 378, 380, 381
– humans vs. dummies, 241 – – compression, 377, 378
– Hybrid III, 253, 254 – – general, 375
– legal crash dummies, 239, 253 – voxel models, 250
– new generation/SIMon, 259 fibula, 342, 343, 349, 375, 376
– SID, EuroSID, BioSID, SID II(s), 256 Fick’s law, 132
– side impact dummy models, 257 field variable interpolation, 159
– THOR advanced 50th percentile male, 258 finite differences, 171
dura mater, 246, 307, 412 finite element method(s), 151–153, 155, 158,
Dyna3D, 238 170, 174, 238, 363, 364, 552, 563
finite element model, 596, 597, 601, 641
EFG (element-free Galerkin), 365 finite elements, 364, 366, 596, 598, 642
elastic modulus, 591, 605, 606, 621 – beams and bars, 363
elastic properties, 596, 601, 606, 609, 651 – joints, 363
662 Subject Index
– membranes, 313, 337, 363, 367 geometry, 596, 599, 605, 608, 623, 641, 651
– plates and shells, 363, 365 geometry acquisition, 245, 288, 290, 297
– solids, 292, 363, 365, 431 geometry conservation law (GCL), 85
– springs, 363 glandular tissue, 620, 622, 638
– thin shell, 354, 365, 366 global degrees of freedom, 603
flail chest, 329 Goldman–Hodgkin–Katz equation, 132
flesh Green integration formula, 36
– lower extremity, 346 Gronwall’s lemma, 37, 199, 210, 211, 216
– meshing, 428
– models, 292 H-Ankle&Foot
– padding, 337, 345 – anatomy, 349
– upper extremity, 339 – calibration
fluid particle, 19 – – bone, 352
fluid–structure interaction (FSI), 74, 421, 424 – – ligaments, 354
– iterative algorithm, 80, 85 – injury, 350
– one-dimensional models, 87 – model, 351
focal length, 164 – validation
foot well intrusion, 351 – – dynamic load, 355
force feedback, 642 H-ARB
force–calcium relation, 139 – calibration, 299
force–velocity relation, 141, 142 – model, 299
forward problem of electrophysiology, 171
– validation and performance, 301
FPM (finite point methods), 365
– with neck models, 316
fracture
– with UE models, 338
– acetabulum, 343–345
H-Dummy, 298
– ankle/foot, 350
H-Head
– hip, pelvis, 343, 345
– anatomy of skull and brain, 302
– lower extremity, 343, 344
– injury, 306
– malleolar or bimalleolar, 350
– mechanical properties, 308
– neck, 314
– validation, 308
– rib fracture, 243, 292, 332, 333, 335, 358, 359
H-LE
– skull fracture, 243, 306
– anatomy, 342
– talar, 350
– tibial pylon, 350 – calibration, 345
– upper extremity, 335, 337–339 – injury, 343, 344
free boundary, 222, 226 – model, 345
free vibration, 368, 369 – validation
friction parameter, 92 – – frontal crash, 345
frontal pendulum impact (thorax), 328, 329 – – side impact, 348
FSI (fluid–structure interaction), 363, 365, 421, H-Neck
425 – calibration, 313
fura, 140 – injury, 313, 314
fuzzy connectedness, 601 – model, 313
– model for whiplash, 276, 277
gadolinium pills, 624 – rear impact with headrest, 319, 320
Galerkin, 196, 198, 200 – validation, 316
Galerkin procedure, 155 H-Torso
gap junctions, 133, 150, 166 – anatomy, 321
Gauss theorem, 149, 150 – calibration of ribs, 327
Gaussian quadrature, 174 – injury
gel system, 651, 652 – – spine, 325
generalised Stokes, 46 – validation, 333
generalized derivatives, 159 H-UE
Generic Programming, 173 – anatomy, 335
Subject Index 663
– calibration, 337–339 – MADYMO model, 268

– injury, 337 – meshing process, 291
– model, 337 – Radioss model, 269
– validation, 340, 341 – rib model (calibration–validation), 292
haptic feedback, 454, 457, 459–465, 478, 510, – thorax model (calibration–validation), 293
511, 539, 542 – Validation Data Base, 251
HARB Huxley’s crossbridge theory, 142
– applications, 281 Hybrid III
– joint modeling, 273 – 5th percentile female model, 357, 358
– model families, 280 – chest belt loading, 333
– models, 271, 272 – Hybrid III dummy family, 253
– passive muscle modeling, 274 – models, 254, 255, 267
– validation, 283 – sled test comparison, 301
Head Injury Coefficient (HIC), 255, 256, 260 hyperelastic, 609, 612
head mask, 322 hypertrophic cardiomyopathy, 166
headrest, 317, 320, 321
health care, 596 iliac wing, 345
health care cost, 350 ilium, 342
heart Image Acquisition, 601
– general, 242 imaging modalities, 597
– modeling, 357, 423
imaging-guided localization, 594
Heaviside function, 192, 193, 195
immune system, 191, 194, 195
heterogeneity, 146, 147
implicit integration, 369
hexahedral element, 649–651
implicit methods, 364
HIC (Head Injury Coefficient), 255, 256, 260
implicit scheme, 222, 224
high-field, 596
implicit surfaces, 599
Hilbert space, 197
impulsive pressure wave, 116
Hill coefficient, 141
incompressible flow, 26
Hill model
inertial forces, 591, 603
– activation function Na , 418
infraspinatus, 335
– active force, 414, 415
– neuro-control flag, 417 inhomogeneity, 638
– neurological reflex time, 417 injury
– optimum muscle length, 415 – abdomen, 326
– passive force, 416 – coefficients, 255
– physiological cross section area, 275, 400, – criteria, 255
414, 418 – foot/ankle complex, 350
Hill Muscle Model, 413 – head, 306
Hill’s equation – lower extremity, 330
– force–velocity relation, 141–143 – neck, 313, 314
– sigmoidal response curve, 143 – spine, 325
hip dislocations, 343, 345 – thoraco-lumbar spine, 324
Hölder inequality, 215 – thorax, 243
hollow organs, 242, 295, 421, 422, 425 – upper extremity, 335, 336
Hong-Ik University, 271, 297 INRETS, 251, 252, 265, 297
hourglass modes, 365 – LBA, 297
human volunteers, 239–241, 251, 253, 301, 302, – LBMC, 297
311, 316, 322, 419 Instron, 609
HUMOS integrated modeling, 130, 138, 174
– biomaterial properties, 293 integration error, 604, 638
– biomaterial response curves, 248 inter-vertebral discs, 292, see also: discs
– geometry acquisition, 288 – mechanical properties, 399
– HUMOS-1 Project, 287 interactions
– interaction between parts, 427 – contact simulation, 427
664 Subject Index
– mesh merging, 428 ions, 131, 137

– sliding interfaces, 429, 430 – calcium, 132, 136, 138, 140, 142, 143
– tied contact, 427 – potassium, 132, 134, 136
interactions between parts, 427 – sodium, 131
intercalated discs, 166 IPS International (IPSI), 271, 297, 298
internal organs ischium, 342
– intestines, 242, 295, 324 isometric contraction, 139, 142
– kidneys, 246, 409 isometric tension, 139–141
– liver, 407 isoparametric, 601, 640, 645
– lungs, 407 isoparametric element interpolation, 156
– pancreas, 324 isoparametric mapping, 156
– spleen, 407 isotonic force–velocity relation, 141
– stomach, 242, 246, 324, 424 isotropic material, 643
interpolation functions, 158, 159, see also: basis
functions Jacobian, 645
– bicubic Hermite, 165, 179 joint finite elements, 299
– cubic Hermite, 176 joints
– Hermite, 159, 175 – acromio-clavicular joint, 274, 335, 336
– Lagrange, 159, 175 – ankle, 349
– linear Lagrange, 175, 178 – Chopart, 350
– quadratic Lagrange, 175, 178 – facet joints, 323, 325
– tricubic Hermite, 180 – gleno-humeral, 274, 335, 337
– trilinear Lagrange, 180
– hip joint, 342, 343, 345, 389
intestines, 242, 295, 296, 324, 326
– knee, 273, 342, 343, 345
inverse Jacobian, 645
– Lisfranc, 350
inversion and eversion, 355
– lower limb joint, 397
ion channels, 132, 133, 135–139, 147, 148, 172
– – ankle, 397
– active, 132
– – hip, 397
– passive, 132
– – knee, 397
ion concentration, 132, 137, 139, 146
– mechanical properties, 389, 395, 398
ion currents, 132–134, 136, 138, 170
– modelling, 273, 299, 312, 337
– calcium, 137
– skeletal (Robby), 272
– calcium activated chloride, 138
– upper limb joint, 396
– delayed rectifier potassium, 138
– – elbow, 396
– fast inward sodium current, 137
– L-type calcium, 137, 138 – – shoulder, 396
– potassium, 137, 147 – – wrist, 396
– sarcoplasmic, 138 junctional sarcoplasmic reticulum, 138
– sarcoplasmic pump, 137
– sodium, 137, 139 kidneys, 246, 409
– sodium–calcium exchanger, 137 Kirchhoff–Love plate theory, 365
– sodium–potassium pump, 137 knee bolster, 238, 241, 298, 345, 347
– T-type calcium, 137, 138 knee padding, 343, 345
– transient outward potassium, 138 Korn inequality, 37
ionic models, 130, 135, 139, 172–174
– Beeler–Reuter, 136 LAB, 265, 267, 296
– Bernus, 139 Lagrange equations, 603, 604
– Fitz–Hugh–Nagumo, 138 Lagrange formulation, 596, 606
– Hodgkin–Huxley, 133–136 Lagrangian, 363, 366, 424
– Luo–Rudy, 136–139 Lagrangian derivative, 21
– Priebe–Beuckelmann, 139 Lagrangian variables, 19
– Puglisi–Bers, 138 landmarks, 615, 618, 624, 625
– species specific, 138 large arteries, 10
– Winslow, 138 large deformations, 596, 606, 608
Subject Index 665
lateral pendulum impacts, 329 – interoperative, 568

latissimus dorsi, 278, 335 – phase contrast, 576
law of Hooke, 481, 531 – tagging, 576
law of mass action, 134 major novelties, 641
least squares, 152–154, 162–165, 167, 169 mapping
lemma of Gronwall, 37, 199, 210, 211, 216 – ALE, 73, 81, 84
length transient, 141 – Lagrangian, 19, 72
length–history dependence, 141, 142 Markov Random Field, 561
length–tension relation, 139–142 Marseille University, 297
length-dependent activation, 142 mass density, 603
lesion, 594, 601, 617, 620, 625, 632–634, 636, mass matrix, 366
642 mass–spring models, 599
levator anguli, 335 master segment (contact), 372, 373
linear elastic model, 555, 572 material derivative, 21
– transversely isotropic, 556 material law, 248, 292, 297
linear elasticity, 58, 474–476, 478–483, material model, 555
487–491, 494–496 – linear elastic, 555
– isotropic material, 480, 481, 487–491 material properties, 596, 597, 601, 603, 605,
– transversal anisotropy, 481, 482, 494–496 612, 620, 623, 625, 630, 637, 638, 640,
linear system solvers, 173 642
linear tetrahedron element, 476, 477, 483–486, maximal monotone graph, 192, 195
491–494, 497, 498, 501, 514 mean velocity, 88
linearization, 364 Mecalog, 262, 269
Lipschitz continuous, 200, 207
mechanical testing, 609
liver, 404, 408–410, 431
mechanoelectric feedback, 130, 174
– anatomy and physiology, 467–474, 476–478,
medical images, 551
499, 504, 517, 528, 530, 538, 539, 542
medium arteries, 10
– injury, 326
membrane potential, 131–133, 135, 136, 172
– mechanical properties, 407, 409, 410, 412
meniscus (lateral, medial), 399
– model, 328
mesh generation, 571
load balancing, 172
– left ventricular, 578
load cell, 609
mesh refinement, 365
load vector, 366
mesh-less methods, 363, 365
lobular carcinoma, 620
longitudinal tubules, 138 meshing process, 291
lower extremity Mindlin–Reissner plate theory, 365
– H-LE model, 341, 342 minimizing subsequence, 219
– Madymo model, 267 misclassification, 625, 629, 630, 638
– meshing, 428 mitosis, 190, 192, 193
– Radioss model, 269 Model Dynamics, 603
– validation, 345 model geometry, 596
lumped mass matrix, 364 model performance, 623–625, 636, 640
lungs, 407 modeling errors, 639
– airbag model, 293, 422 moment–angle relationships, 301, 389
– anatomy, 323 momentum equation, 29
– injuries, 326 momentum–flux correction coefficient, 88
– mechanical properties, 407 Mooney–Rivlin, 609, 611, 612
– model, 328 MRI, 591–593, 622, 625, 630
multi-body models, 237, 254, 255, 268, 271,
M cells, 132, 146 275, 283, 298, 302, 372
macro modeling, 250 multicoil array, 594
MADYMO, 268 Murnaghan equation of state, 307
magnetic resonance imaging muscle bars
– cardiac, 577 – arm–shoulder model, 274
666 Subject Index
– Hill muscle bars, 276, 277, 413 optical mapping, 146

– Hill-type muscle bars, 275, 418 optimization, 222
– lower extremity (LE), 337, 345
– static force distribution, 276, 279 pacemaking activity, 144, 145
muscle tonus, 296 Pam–Crash, 238
muscles pancreas, 324, 326
– actine tubes, 417 papillary muscle, 163, 166
– age differences (rectus abdominis), 404 parallel programming, 172, 174
– myosin fibers, 417 parametric models, 599
– sarcomere cells, 417 parenchyma, 601, 603
– tensile strength partial activation, 142
– – lower extremity, 403 particle methods, 363, 425
– – trunk, 403 passive muscle action
– – upper extremity, 403 – arm–shoulder model, 274
myocardial infarction, 575 – H-ARB neck, 301
myofiber disarray, 166 – Hill model, 276, 413
myosin, 142–144 – lower extremity, 337, 345
– mechanical properties, 275, 400, 418
Navier–Stokes equations, 31 passive safety systems, 254
– Chorin–Temam scheme, 54
passive transport, 132
– energy inequality, 41
patch clamping, 135
– factorisation schemes, 51
patella, 343
– finite element discretisation, 46
patella fracture, 343
– projection methods, 49
PATRAN, 598, 608, 612
– weak form, 38
pectoralis
– Yosida scheme, 54
– major, 335
NCAP, 253
– minor, 335
necrosis, 190, 195
penalty contact, 372
needle, 594, 595, 597, 641, 642
penetration (contact), 265, 372, 429
Nernst potential, 132, 133
perforation (contact), 373
network sarcoplasmic reticulum, 138
neurosurgery, 551 pericardium, 163
Newtonian fluid, 30 phantom construction, 608, 651
NHTSA, 239, 253, 256, 258–261 phantom studies, 599
nodal constraints, 372 physical models, 598
nodal displacement, 604, 605, 646–648 physical slicing, 288
nodal force, 605, 649 physics-based, 599
non-linear elasticity, 476, 530–542 pia mater, 305
non-rigid motion, 551 piecewise continuous, 196
numerical integration, 605, 638 plantarflexion, 351
numerical simulation, 189, 190, 226 plasma, 10
plate compression, 617, 626, 628
Object-Oriented Programming, 173 platelets (thrombocytes), 10
occupant surrogates (see dummies), 237, 240, PMHS (Post Mortem Human Subject), 239, 240,
245, 253 299
Ohm’s law, 133, 147, 149, 150 Poincaré inequality
1D models of blood flow, 87 – multidimensional case, 37
– conservation form, 94 – one-dimensional case, 68
– energy conservation, 100 point collocation, 153
– energy inequality, 103 Poiseuille flow, 12, 88
– hyperbolic system, 96 Poisson problem, 51
– quasi-linear form, 94 Poisson ratio, 63, 643, 647, 649
– weak solution, 105 polynomials, 156, 159
operator splitting methods, 171 – Hermite, 159
Subject Index 667
– Lagrange, 159 – calibration, 327

posterior cruciate ligament (PCL), 343 – fracture, 243, 292, 327, 359
potential gradient, 132 – injury, 326
precomputed linear elastic model, 454, 462, 463, – mechanical properties, 376
465, 466, 477, 478, 502–512, 519, 524, – modelling, 264, 291
542, 545 – validation, 292, 332
pressure, 30 Riemann invariants, 97
pressure feedback, 642 rigid walls, 372
pressure–area relationship, 93, 101, 103 RMDM (Relative Motion Damage Measure),
pressure–volume relation, 140 260
principal direction, 602 rotational differences, 635, 642
principal stretches, 610 ryanodine, 138, 140
principle
– of conservation of mass, 25 sacrum, 323, 342, 395
– of conservation of momentum, 26, 28 sagittal plane and section, 271, 272
pronation, 351, 357 sarcolemma, 140
proximal, 14 sarcomere, 139, 141, 174
pseudo-characteristic variables, 99 – length, 139–141, 143, 144
pubic rami, 345 – length–history, 142
Purkinje fiber network, 145, 169 sarcoplasmic reticulum, 138, 140
PVC, 652
scale factor, 158, 159
Python, 173
scaling, morphing, aging, 245
scalp, 306, 307
quadriceps, 343
scapula, 242, 274, 292, 335, 375
SDH (see brain injury), 306
Radioss, 269, 296
segment validation, 292
reaction terms, 196, 203
segmentation, 601, 642
reaction–diffusion equations, 131
real-time, 599, 642 – brain, 569
rear impact, 240, 245, 283, 285, 301, 302, 316, – left ventricular, 578
319, 320, 400 sensitivity analysis, 598, 623, 625, 630, 637,
recirculation, 15 638, 640
red cells (erythrocytes), 10 serratus, 335
references shaft fracture, 343, 345
– bio-materials (further), 412 shape functions, 643
– biomaterials, 251 shear strains, ventricular, 168
– H-UE model, 340 shortening velocity, 141, 142
– injury and trauma, 243 shoulder, 240, 267, 274, 298, 322, 334, 335,
– LAB-Model, 267 337, 338, 340, 341
– multi-body and muscle models, 283 SID, 240, 253, 256, 257
– muscle materials, 419 side impact, 237, 240, 241, 298, 311, 335, 348
– SIMon-Head, 262 silicon gel phantom, 597, 608, 609, 652
registration, 597 SIMon, 259–262
relative refractory period, 139 – Damage Measure
relaxation method, 604 – – CSDM, 260
relaxation-based elastic models, 522–524 – – DDM, 260
reliability, 598, 637, 640 – – RMDM, 260
repolarization, 132, 133, 138, 146, 147 simulation of cutting, 454, 455, 461, 463,
residual, 152, 154, 174 465–467, 475, 476, 502, 512–514,
resin, 651 517–519, 521, 522, 524, 525, 529, 535,
Reynolds number, 43 538, 542
rhomboids, 335 simulation of fracture, 365
ribs simulations of deformation, 597, 599
– anatomy, 321 sinoatrial node, 145
668 Subject Index
skeletal muscle, 139, 141, 142 strain energy, 609, 610

skin, 602, 604, 617, 618, 620–622, 624, 648 strain rate tensor, 30
– mechanical properties, 404, 407 strain tensor, 465, 477, 478, 482, 484
– thickness, 620 – finite strain, 530–532, 536
skull – infinitesimal strain, 478–481, 483, 488, 491,
– cross section, 248 530
– diploe, 303 strain-rate sensitivity, 292
– fractures, 306 stress, 606, 607, 609–611, 620, 621, 638, 643
– H-Model, 306, 307 stress–strain, 607, 609–611, 620–622, 638, 643
– HUMOS model, 291 stroke work, 140
– mechanical properties, 308, 376 strong formulation, 152
– outer table, 303 structure models, 59
– SIMon model, 261 – algebraic model, 65
– vitreous/inner table, 247, 303 – analysis, 68
slab model, 625, 628–630, 640 – generalised string model, 66
slave node (contact), 372 – – viscoelastic term, 79
sled test, 252, 253, 281, 282, 284, 297, 301, 309 – independent ring model, 63, 74, 108
slicing, 245, 299 subarachnoidal space, 306
sliding filament theory, 142 subclavius, 335
small strain theory, 596, 606, 608 subscapularis, 335
smooth particle hydrodynamics (SPH), 363 subspace, 138
solid organs, 242, 244, 326, 327 summation property of integrals, 155
solution method, 363 superior vena cava, 145
solution stability, 364, 368
supination, 351, 357
sources of error, 638, 639
supraspinatus, 335, 404, 405
SPH (smooth particle hydrodynamics), 363, 365,
surgery planning, 597
425
surgical navigation, 567
spinal chord, 323
symmetry, 192, 193, 209, 222
spinal column, 292, 321, 325, 342, 399
systole, 12
spine
– cervical, 243, 292, 312–314, 320, 329
– lumbar, 243, 323, 324, 326, 327, 334, 347, 398 T-wave, 146
– thoracic, 321, 323, 324, 334, 335, 398 tectorial ligaments, 315
spleen, 242, 246, 256, 324, 326, 407–409 tele-surgery, 597
spongious bone, 248 tendon
spring–mass models, 465, 466, 475, 499, 521, – Achilles, 347, 350, 405
522 – finger flexor, 405
stability – patellar, 405
– explicit solution, 368, 369 – supraspinatus, 405
stable time step, 368 tensor–mass models, 465, 466, 513–522, 524,
Stapp, 239, 240 526–528, 533, 545
Starling’s law, 144 tentorium cerebelli, 306
stenosis, 15 teres, 335
step size, 604, 605 tetanic contraction, 139, 142
sternum, 242, 274, 292, 293, 324, 326, 375 theorem
stiff ODE solvers, 173 – ALE transport, 73, 89
stiffness, 604, 605, 643, 648 – Cauchy stress tensor, 28
stiffness matrix, 366, 370, 377, 604, 605, 643, – Helmholtz decomposition principle, 50
648 – Reynolds transport, 24
Stokes equations, 43 – Sobolev embeddings, 36
Stokes pressure matrix, 52 theory
stomach, 242, 246, 324, 326, 327, 422 – crash codes, 364
strain, 591, 596, 600, 606–608, 610–612, 614, therapy, 191
620, 621, 623, 638, 643, 644 θ -rule, 171
Subject Index 669
thick filaments, 139 – brain models, 308

thin filaments, 139 – female H-Model
thorax – – impact test, 357
– anatomy, 323 – Female H-Model
– female – – out-of-position airbag inflation, 360
– – calibration, 357 – female thorax, 361
– – model, 357 – H-Ankle&Foot
– – validation, 361 – – plantar impacts, 355
– H-Model, 326 – – static crush, 355
– HUMOS model, 291 – H-ARB, 301
– injury, 243 – H-Head, 308
– NHTSA model, 261 – H-LE, 347, 348
– validation HARB model, 283 – H-Neck, 316
– validation HUMOS model, 293 – H-Torso, 333
threshold voltage, 131, 132 – H-UE, 340, 341
tibia, 342, 343, 377, 378, 397 – HARB, 283
time discretization, 364 – lower extremity, 345
time integration schemes, 172 – ribs, 332
– backward Euler, 171 – thorax, 328
– Crank–Nicolson, 171 vascular prosthesis model, 115
– forward Euler, 171 vascularization, 189
time-discretized nodal equation, 604, 638 velocity, 20
tissue
venous sinuses, 306
– material properties, 411
ventricles, 306, 307
TNO, 268, 269, 297
ventricular activation, 145
torso, 257, 278, 283, 298, 304, 313, 324
ventricular architecture, 145
Total Human Model for Safety (THUMS), 262,
– fibers, 133, 145, 146, 162, 166–170
264, 266
– orthotropic, 168
trabecular bone characteristics, 353
– sheets, 133, 145, 146, 162, 168–170
trabeculata–compacta interface, 166
– transversely isotropic, 168
traction stress, 66
ventricular geometry, 163
translational differences, 635
ventricular microstructure, 145, 146, 162
transport formula, 25
transverse plane and section, 271, 272 vents, 295
trapezius, 275, 335 vertebrae, 291, 292, 301, 311, 313, 321,
triceps, 335 323–325, 395
trilinear, 601 VIEWPOINT, 271, 280, 281, 283, 299, 313
trochanter, 343 virtual deformable breast model, 596
tropomyosin, 143 virtual surgery, 597
troponin C, 138, 140, 142, 143 viscera, 327
truncated ellipsoidal model, 163, 164 visco-elastic, 242, 247, 248, 292, 311, 315, 341,
Tumor Angiogenesis Factor (TAF), 190, 194 400, 409, 410, 413
turbulence, 44 viscoelasticity, 141
viscosity
unconditionally stable, 364, 370 – dynamic, 30
uniaxial deformation, 610 – kinematic, 31
uniqueness, 190, 203, 205, 208, 214, 216, 220 viscous forces, 141
unit vectors, 159 visual rendering, 454, 457, 459–465, 476, 478,
University of Heidelberg, 283, 293 504, 510–512, 516, 517, 519, 525, 526,
University of West Bohemia, 271, 275, 327 529, 539, 542
unstructured meshes, 366 vitamin E pills, 597, 615, 617, 618, 624, 627,
640, 641
validation voltage-gating, 132, 134
– abdomen, 295 volume calculations, 649
670 Subject Index
volume conservation, 649 weight functions, 153–155

volume elements, 601, 602, 612 weighted residuals, 152–154, 169, 170
volume ratio, 610 white cells (leukocytes), 10
volunteer, 252, 283, 301, 322 Womersley flow, 12
von Mises stresses, 332, 333, 356
voxel models, 248, 250 X-ray, 240, 245, 250, 312, 323, 336, 342
wall laws, 108 Young modulus, 93, 105, 116

Wayne State University (WSU), 262, 263, 340 Young’s inequality, 199, 218
weak formulation, 153, 156
weak solution, 197, 198, 201, 202, 216, 224 zero-energy modes, 365, 366

Computational Models For The Human Body PDF

Încărcat de

Informații document

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Computational Models For The Human Body PDF

Încărcat de

Drepturi de autor:

Formate disponibile

General Preface

It will also cover the numerical solution of actual problems of contempo-

S PECIAL VOLUME : C OMPUTATIONAL M ODELS FOR THE H UMAN B ODY

S UBJECT I NDEX 657

Alfio Quarteroni a,b, Luca Formaggia b

E-mail addresses: alfio.quarteroni@epfl.ch (A. Quarteroni),

5. The derivation of the equations for the flow field 17

10. The incompressible Navier–Stokes equations and their approximation 35

15. Mathematical modelling of the vessel wall 57

18. The coupled fluid structure problem 71

20. One-dimensional models of blood flow in arteries 87

C HAPTER VII 111

21. Some numerical results 111

2. A brief description of the human vascular system

F IG . 2.2. The human heart. Courtesy of the Texas Heart®Institute.

3. The main variables for the mathematical description of blood flow

4. Some relevant issues

F IG . 4.1. A typical flow rate in an artery during the cardiac cycle.

the propagation of pulse pressure waves. Indeed, no propagative phenomena would

5. The derivation of the equations for the flow field

F IG . 6.1. Example of admissible domains. Ωa has a boundary formed by piece-wise C 1 curves. Ωb is a

Given a function f : Ω → R, x → f (x), and a domain V ⊂ Ω, we will use the

7. The motion of continuous media

F IG . 7.1. The Lagrangian mapping.

numerical approximation of the problem at hand it will be necessary to introduce yet

7.1. The velocity

7.2. The material derivative

7.3. The acceleration

In the Lagrangian frame the acceleration is a vector field â : I × Ω0 → R3 defined as

7.4. The deformation gradient

In particular, its determinant,

T HEOREM 7.1. Let Vt ⊂ Ωt be a subdomain of Ωt and let us consider the function

where fˆ(t, ξ ) = f (t, Lt (ξ )). In short,

7.5. The Reynolds transport theorem

P ROOF. We have, by direct application of the chain-rule,

We now exploit the well-known result that for any matrix A,

det(I + εA) = 1 + ε tr A + o(ε),

Jt +ε = det(  Jt + o(ε) = (1 + ε div û)Jt + o(ε).

We have now the following fundamental result.

T HEOREM 7.2 (Reynolds transport theorem). Let V0 ⊂ Ω0 , and Vt ⊂ Ωt be its image

Furthermore, we exploit again the definition of material derivative (7.3) in order to

The second equality in (7.11) is a consequence of (7.4).

8. The derivation of the basic equations of fluid mechanics

8.1. Continuity equation or mass conservation

We assume that there exists a strictly positive, measurable function ρ : I × Ωt → R,

constant density fluid ⇒ incompressible flow,

8.2. The momentum equation

Another important principle allows the derivation of an additional set of differential

Three different types of forces may be acting on the material inside Ωt :

force acting on a volume Vt is given by

8.2.1. The Cauchy principle

R EMARK 8.2. In deriving (8.9), we have assumed that Vt is completely contained

∂ui ∂ui ∂Tij

The non-linear term ρ(u·∇)u is called the convective term.

Indeed, thanks to the incompressibility of the fluid,

(u·∇)u = (u·∇)u + u div u = div(u ⊗ u).

8.3. The constitutive law

Since ρ is constant, it is sometimes convenient to introduce the kinematic viscosity

9. The Navier–Stokes equations

u(t = t0 , x) = u0 (x), x ∈ Ω. (9.2)

9.1. Boundary conditions for the Navier–Stokes equations

The Navier–Stokes equations must be supplemented by proper boundary conditions that

T · n = −P n + 2µD(u) · n = te on Γ n ⊂ ∂Ω, (9.3)

Jt +ε = det( Jt + o(ε) = (1 + ε div û)Jt + o(ε).

f2H1 (Ω) = f2L2 (Ω) + ∇ f2L2 (Ω) .

L EMMA 11.2. Let Ω be a domain of RN and let L ∈ V . Then L(v) = 0, ∀v ∈ Vdiv if

P ROOF. We first prove the non-singularity of C. For any W ∈ RNVh , W = 0,